Repository navigation
Releases
Anti-REA Forge v1.4.0
Anti-REA Forge 1.4.0 — Complete-content protection and clearer workflows
Release date: 2026-11-18. Python core and JavaScript SDK use 1.4.0.
The existing v1 signed-document formats, CLI/import names and default JSON protocol
remain compatible. This release adds no runtime dependency to the SDK.
Version 1.4 closes two publication gaps reproduced against the official v1.3.0
wheel and makes protection references easier to configure and diagnose. A release
gate controls what is shipped; distributed code and disclosed keys remain readable
by REA and other independent consumers.
Measured improvements
| Control | Official v1.3.0 | Version 1.4.0 | Evidence |
|---|---|---|---|
| Complete private source with a prefix, suffix or concatenated bundle | Allows all three wrappers | Positively rejects all three | Both real REA distributions identify the owned function and its constants; a recovered module produces the same four results |
| Locating value crossing a 128 KiB window | Allows the containing file | Rejects with a completed content check | Separate position control, not an additional vulnerability |
| Wrapped content in six ZIP/ASAR layouts | Allows all six deployments | Rejects all six | Disk, nested members and verified unpacked siblings share the production decision; layouts are not added to REA recovery counts |
| Six supported keys inside JSON strings | Allows all six usable keys | Positively detects all six | Real Node JSON.parse plus Ed25519 signatures or AES-256-GCM decryption |
| Six recoverable over-budget or ambiguous JSON inputs | Allows all six | Refuses all six as incomplete inspection | Kept separate from positive key findings |
| Six normal structured-data controls | Allows all six | Allows all six | Includes public JWK, ordinary text, a 1.1 MiB string and 260 empty container placeholders |
| Initialize with private/key references | Flags are unavailable | Validates and saves configuration-relative paths | Zero manual configuration edits; a moved project works through the real CLI and Node SDK |
| Explain one exact-copy finding | 74 lines / 4,695 characters of JSON | 23 lines / 1,026 characters of text, with a next action | One fixed report: 68.92% fewer lines; not a scanning-speed or general usability score |
The content controls use npm REA 6.9.0 and a separately built source tree at
675287517c98c702ee88ad036c5e3b8dd4bcc7a7. These are actual deterministic CLI
providers, not an autonomous model session. Key and workflow controls use real
standard consumers and do not invoke REA. All fixtures are owned, synthetic inputs.
Records: content / npm,
content / source,
stringified keys,
workflow,
full assessment.
Complete private contents inside larger payloads
Existing private_input_files references now use exact-content/v1 automatically.
The complete unchanged bytes can appear anywhere inside an inspected physical file,
ZIP/ASAR member or validated unpacked sibling. Adding a build header, trailing newline
or concatenated public module no longer avoids the check.
References retain their limits: at most 64 files, 1 byte to 8 MiB each, 64 MiB total.
A prefix of at most 64 bytes locates a long candidate, then its entire original length
is verified with SHA-256. Short references match their complete bytes, including
one-byte references; common short values can therefore match ordinary public content.
Duplicate references share work while preserving rule indices.
The whole release shares 4,096 candidate checks / 512 MiB of candidate verification
reads, in addition to existing release/container limits. Exhaustion produces
private_input_unverifiable and private_input_scan_completed: false; it never
becomes a successful absence result. New reports include the profile, checked payload
count, candidate count and charged verification bytes. Contained positives retain
private_input_exposure and add representation: "contained-bytes".
Snapshots and reports do not expose external reference paths, locating bytes, contents
or separate reference hashes. Public file/member paths and public inventory hashes
remain visible. prepare retains its scan-to-sign input rechecks and rollback on a
failed final verification. See full profile.
Stringified JSON keys with shared limits
The structured-secret scanner parses at most two additional JSON-string container
layers along a nested path, then uses the existing private JWK, PEM/DER and known-key
rules. Each container is limited to 64 KiB UTF-8 text. All branches and DER/PEM
decodes share 64 candidates / 256 KiB, and all parsed nodes count toward the
100,000-node / 64-depth limits.
Recognizable overdeep/oversized containers, duplicate members and nonfinite numbers
fail as incomplete inspection. Ordinary text and obvious templates skip the parser.
Syntactically empty objects, arrays and empty quoted strings need no parse and consume
no decode budget. Nonempty public containers still count: 65 serialized public objects
can exceed this bounded profile even without private material.
Positive nested findings have a fixed json-stringified-... representation and
stringified_layers count. Decoded values, field names and key digests are absent
from the findings. No JavaScript evaluation or arbitrary encoding recursion is added.
Configure once and understand failures
Use matching 1.4.0 core and SDK archives from this release:
python -m pip install --upgrade https://github.com/Afloat16/anti-rea-forge/releases/download/v1.4.0/open_release_guard-1.4.0-py3-none-any.whl
npm install https://github.com/Afloat16/anti-rea-forge/releases/download/v1.4.0/anti-rea-forge-releaseguard-1.4.0.tgzFor a new configuration, supply the protection inputs that exist in your project:
python -m releaseguard project-init --root dist --product MyApp --out .releaseguard \
--private-input-file service/private_quote.py --asset-key-file issuer/asset.key
python -m releaseguard check --config .releaseguard/project.json --format textBoth reference flags are repeatable. Omit whichever kind your product does not use.
Inputs must already exist; dist may be built later. Initialization validates inputs
before writing, saves paths relative to project.json, rechecks them after writing,
and removes a new configuration bundle if final validation fails. Existing config
directories are not overwritten. Existing configurations require no schema migration;
edit their private build configuration only when adding references.
scan-release, check and prepare accept --format text. Reports show configured
reference counts, enabled/completed inspections, public locations, rules and fixed
remediation actions. Zero references are explicitly marked not configured.
Dynamic strings use escaped JSON-style quoting, including control and Bidi characters.
JSON remains the default, and --format json is explicit. Exit codes remain 0
(configured checks pass), 3 (publication blocked by findings), 2 (operation
error) and 130 (cancelled). Failures retain stdout/stderr separation. prepare
sidecars remain JSON with the existing signature format. The SDK keeps the machine
protocol and does not acquire a signing or authorization-issuance method.
Validation and publication
Local validation ran 1,066 Python test cases: 1,065 passed and 1 skipped, plus
10 test classes skipped during setup for unavailable platform/compiler dependencies.
The unittest summary therefore contains 11 skip records; setup-class skips are not
subtracted again from the executed-case total. The SDK passed 67 tests and 12 real
CLI integrations. The change adds 65 Python test methods. Skips are recorded and
not counted as successes. Validation record
The required CI set remains 13 jobs across Linux, Windows and macOS, including
real LLVM 18, Ghidra, REA npm/source, official ASAR, native compilers and fresh package
installations. New content, key and workflow controls are mandatory in the existing
REA job. Fresh wheel/sdist must enforce containment, stringified-key rejection, saved
references and readable reports; the packed SDK must enforce containment and saved
references. The publisher requires each relevant installation flag to be strictly true.
Public assets include the wheel, source archive, SDK archive, SHA256SUMS,
installation-report.json and release-evidence.json. Publication binds successful
CI and installed artifact checks to the exact main commit and archive hashes.
Remaining boundaries
The changed-comment source fixture deliberately remains allowed and REA-recoverable:
complete unchanged reference bytes are absent even though its function behaves the
same. Rewrites, minification, compilation, encoding, fragments, undeclared private
files and arbitrary application secrets remain outside the content profile. Keep
private service implementations out of public build inputs; this scanner does not
automatically partition an application.
The matching resource probe is a local bounded-work measurement, not a total process
memory or general throughput guarantee. Trusted build paths, stable staging files,
external keys and independently enforced service authorization remain required.
Existing historical recovery failures remain published. The release does not claim
that REA, another file reader or an authorized runtime can no longer analyze software.
Anti-REA Forge v1.3.0
Anti-REA Forge 1.3.0 — Private inputs and residual disclosure channels
Release date: 2026-11-18. Python core and JavaScript SDK use 1.3.0.
The existing v1 signed-document formats and CLI/import names are unchanged.
This release closes specific publication gaps reproduced against the official
v1.2.0 wheel. It adds a production interface for current private-file references,
checks configured asset keys in Python bytecode integers, separates eligible
embedded LLVM compiler metadata, and detects private PEM markers inside one
outer JSON hex/Base64 encoding. Ordinary distributed code remains analyzable.
Measured changes
| Control | Official v1.2.0 | v1.3.0 candidate | Required positive / compatibility evidence |
|---|---|---|---|
| 12 canonical Python integer-key representations | Allows all 12; real REA constants reconstruct the exact key | Positively rejects all 12 on disk and in executable ZIP apps | Each recovered key authenticates and decrypts an owned AES-GCM payload; normal PYC remains executable |
| Actual Clang embedded IR after symbol stripping | Gate allows output; llvm-dis recovers 4 private names, a source path and arithmetic |
Reports compiler metadata; splitter removes the eligible public IR channel | Exact private bitcode/command bytes retained; original, old public and new public runtime results are [48, 85, 122]; .text and dynamic symbols unchanged |
| Actual raw and wrapped compiler bitcode | Allows both, renamed and in nested ZIPs | Refuses all 4 layout cases as unsupported inspection | llvm-dis really recovers both compiler-produced inputs; no claim of full bitcode parsing or macOS execution |
| Private implementation accidentally copied into a JS release | Default policy allows the source file | Configured live reference rejects the exact copy | Real REA private-source nodes 1 → 0 → 1 for accidental / public-only / restored layouts; all graphs complete |
| Current private source differs from an old manually entered digest | Old digest permits the changed copy | Live reference rejects the current exact copy | A new snapshot is taken on each check and rechecked through prepare |
| 8 outer PEM representations | Allows all 8; actual Ed25519 signatures verify | Positively identifies all 8 | Two ordinary public/nonsecret controls still pass |
| 2 recoverable PEM representations outside decode budgets | Allows both | Refuses as incomplete inspection | Actual decoding/signing still succeeds; excluded from the eight positive detections |
The Python and private-source controls run against both the pinned npm REA 6.9.0
and the source build at 675287517c98c702ee88ad036c5e3b8dd4bcc7a7. They exercise real
CLI providers, not an autonomous model session. Compiler and PEM controls use
actual standard readers; those experiments do not invoke REA. Fixture counts
are not vulnerability counts or a general resistance/recall percentage.
Baseline integrity
All new paired benchmarks can bind their baseline interpreter to the official
v1.2.0 wheel. Required CI supplies that wheel and verifies its SHA-256:
b16280c1f81cb6cec5a028102edeefdaf8b62e9300da3e87acb1878b3f0ab75b
The experiments compare every installed package file with the verified wheel
and recheck package inventories at completion. Candidate identities and analyzer
versions/build inventories are recorded in the JSON results. A wrong version,
incomplete analyzer output or changed tool/package ends the experiment; it is
never converted into a successful protection result.
New private-file references
Keep the implementation outside the public build and declare its input files
in the build-only project configuration:
{
"version": 1,
"product": "MyApp",
"root": "../dist",
"policy": "policy.json",
"asset_key_files": ["../../issuer/data.key"],
"private_input_files": ["../../service/operation.py", "../../service/config.json"]
}scan-release, check and prepare also accept repeatable
--private-input-file arguments. scanRelease and checkProject in the SDK
accept privateInputFiles. Extra references add to the saved configuration;
an empty extra list cannot disable it. Configured paths resolve relative to the
project file; direct CLI paths resolve from the working directory.
Each reference must be a nonempty regular file outside the release root. The
limits are 64 files, 8 MiB per file and 64 MiB total. No glob expansion,
directory recursion or source execution is performed. Missing files, links,
reparse points, special files and changed inputs fail the operation. The build
operator must control the files and their ancestors. Identity checks and hash
rechecks detect ordinary replacement; they are not an atomic snapshot against
a hostile filesystem owner.
Exact whole-payload copies produce private_input_exposure, including inspected
ZIP and ASAR members and verified unpacked ASAR siblings. Reports show reference
counts and rule indexes without adding external source paths or separate
reference hashes. Existing public inventory hashes remain visible, including
the hash of an accidentally published exact copy.
Transformed, compiled, concatenated, encoded or partial copies do not match
this particular rule. Keep private service code out of the public build graph,
review the allowed output inventory and include all relevant source/config
references. The new service example reuses the
existing run_authorized_operation API. It does not introduce a new HTTP/RPC
protocol or claim a newly discovered authorization bypass. Host authentication,
transport, quotas and access control remain application responsibilities.
Configured keys in Python bytecode
When asset_key_files are configured, recognized stable CPython PYC headers
enter the canonical-marshal-integer/v1 profile. For each known 32-byte key the
scanner derives canonical base-2¹⁵ marshal integer records for big/little-endian,
signed/unsigned reconstruction and their reference-flag variants. It streams
at most 64 MiB and never calls marshal.load, imports examined code, or
executes it.
The separate report fields are pyc_asset_keys_checked,
pyc_asset_key_files_checked, pyc_asset_key_scan_completed and
pyc_asset_key_profile. JSON/DER completion remains independent. A matched
record produces asset_key_exposure with representation
python-marshal-integer; unsupported headers, changed lengths, short integer
aliases or exhausted budgets produce a fixed unverifiable finding.
The header classification covers stable magic values for CPython 3.7–3.14;
the release guard itself still requires Python 3.10+. Header recognition does
not validate every marshal object or execute one version's bytecode under
another version. Each integer interpretation must have at least 128 significant
bits; this is a bounded-pattern condition, not an entropy estimate. An
unconfigured key remains recoverable, and the benchmark preserves that failure
as an explicit boundary. Computed/split constants, noncanonical encodings and
general Python implementation secrecy are outside this profile.
Exact framing and official-source references
Compiler metadata
The native readers report compiler_metadata_present and
compiler_metadata_sections separately from debug evidence:
| Format | Recognized canonical sections | Action |
|---|---|---|
| ELF | .llvmbc, .llvmcmd, .llvm.lto |
Gate finding; eligible data can be split privately |
| PE | .llvmbc, .llvmcmd |
Gate finding; no PE rewriting added |
| Mach-O | __LLVM,__bitcode, __LLVM,__cmdline, __LLVM,__bundle |
Gate finding in each supported slice; no Mach-O rewriting added |
| Standalone bitcode | Raw 42 43 C0 DE and wrapper DE C0 17 0B |
Recognized but unsupported; native metadata completion is false |
scan-release applies these checks under check_native_metadata. The native
gate adds require_no_compiler_metadata, true by default; unavailable ELF
section names cannot establish absence. If a product intentionally ships this
metadata, review and explicitly set the relevant policy opt-out.
The ELF splitter accepts only unique, nonallocated PROGBITS compiler sections
with ordinary flags or SHF_EXCLUDE, no section links/entry semantics, and no
overlap with any file-backed program segment. It checks exact private payload
fingerprints and existing runtime-segment/dynamic-interface invariants. Failure
rolls back its new outputs. Public machine code still implements the algorithm
and remains available to disassemblers.
The checked-in real producer result uses Clang/llvm-dis 19.1.7 on Linux;
required CI independently runs the control with LLVM 18. PE/Mach-O additions
have structural parser regressions, including byte orders and fat slices; those
tests are not new Windows/macOS compiler/runtime measurements.
Wrapped PEM and shared budgets
The JSON scanner recognizes supported private PEM headers inside one outer
hex/Base64 encoding, including leading notes accepted by actual PEM consumers.
It scans bounded hints across the JSON string, then confirms decoded markers.
Node-compatible Base64 handling is reused; a label is conservative disclosure
evidence, not mathematical validation of every possible key.
PEM and DER share a 64-candidate / 256 KiB decoded-byte per-document budget.
Each admitted candidate retains the 256 KiB raw-text / 64 KiB decoded-byte
limit. Even a false alignment hint consumes the budget for its actual decode.
Large unrelated strings remain accepted without private-key parsing; a private
candidate beyond a limit is unverifiable. These budgets sit within the existing
8 MiB / 64-depth / 100,000-node JSON limits. Arbitrary nested wrappers, split keys
and computed material are not evaluated.
Upgrade and release verification
- Install matching 1.3.0 core and SDK archives from the
[release](https://github.com/Afloat16/an...
Anti-REA Forge v1.2.0
Anti-REA Forge 1.2.0 — Expanded recovery controls
Release date: 2026-11-18. Python core and JavaScript SDK use 1.2.0.
This release expands the real REA experiments from single fixtures to different
programs and builds, and closes three demonstrated disclosure paths. It keeps
the previous controls and their failures. The measurements describe what the
specified readers recover from authored release artifacts; they are not a
general probability of preventing reverse engineering.
Measured changes
| Control | Original / previous release | Public / current release | Positive or compatibility control |
|---|---|---|---|
| Five Wasm programs, known private names | 26 recovered | 0 recovered | Private originals restored: 26 recovered |
| Wasm executable behavior | 43 numerical checks, 2 runtime traps, 1 host rejection | Same results; every standard section byte identical | Imports, memory, globals, table mutation and loops exercised |
| Four JavaScript builds, five bundles | 9 embedded originals recovered | 0 exact originals recovered | Private maps restored: all 9 recovered; JS bytes unchanged |
| Incomplete maps with escaped content keys | v1.1 permits 2; REA recovers exact source despite partial graphs | v1.2 rejects both | Complete graph required for absence in the normal three-sided controls |
| Residual Wasm linker symbols | Actual v1.1 split removes name, gate allows, REA recovers one helper |
v1.2 gate and splitter refuse the input | Producer-valid, executable relocatable module; nonexistent outputs are not scored |
| Consumer-compatible key representations | v1.1 permits all 19; recovered material actually decrypts or signs | v1.2 positively identifies and rejects all 19 | Two nonsecret controls pass on both versions |
| Known-key consumer decoding beyond its raw-text budget | v1.1 permits; Node still recovers the exact key | v1.2 explicitly rejects incomplete inspection | Separately reported; excluded from the 19 positive identifications |
Both Wasm and JavaScript corpora run against published rea-agents@6.9.0
and a separately compiled, tracked-clean source snapshot
675287517c98c702ee88ad036c5e3b8dd4bcc7a7.
There are 16 Wasm and 15 JavaScript REA invocations per distribution, including
the separately classified rejection/disclosure controls: 62 actual calls
across the two distributions. Repeated artifact sides and distributions do not
multiply the number of independent programs or behavioral cases.
The source tree and compiled runtime are verified before and after each run.
Official WABT 1.0.42 archive and executable hashes remain pinned. The previous
release comparison uses the official v1.1 wheel with SHA-256
5040a0f33c6b5a640302af4983556abdcab1918e845f4bac4010ec2fa2233086.
Wasm and key probes verify all 29 installed package files against that wheel
before and after execution. The source-map probe's optional baseline check is
version-only; required CI first runs the Wasm identity check on the same
interpreter. Each report states its actual provenance and limits.
1. Finish Wasm linking before release
name is not the only place a helper can be named. A real WABT
--relocatable --debug-names build retains linking and reloc.Code; removing
the debug-name section alone leaves an internal helper recoverable by REA.
The bounded envelope reader now records linker_sections and
evidence.linker_metadata_present. Default portable inspection reports
native_metadata_unverifiable with reason
linker_metadata_requires_final_link and an incomplete metadata flag.
split-wasm-debug rejects before creating either output. The rule recognizes
the section names; it does not pretend to parse linker symbol/relocation tables
or delete data that a linker might require. Some such modules are executable.
The five ordinary modules still separate successfully. Each private original is
byte-exact, all retained public sections keep their bytes and order, and the
public notice custom section survives. Runtime users of removed debug custom
sections can observe the change; test the actual embedding.
2. Treat readable source as a disclosure even in an incomplete map
The escaped JSON key sourcesCont\u0065nt decodes to sourcesContent. Source
text is still readable when mappings is missing, numeric or otherwise not a
valid mapping string. The old structural recognizer required a valid-looking
map and could miss this disclosure while the literal key was already blocked.
The bounded JSON traversal now applies the same conservative key rule after
decoding. It produces source_map_marker with
representation: "json-decoded-key" when existing literal/structural findings
have not already rejected the document. It works within the existing encoding,
XSSI, archive, node and depth bounds. It is a disclosure marker, not a claim
that the document is a valid v3 source map.
The expanded corpus includes three-source builds, two linked bundles, Unicode
sources/keys and a null original. The unavailable original is excluded from
recovery counts. A nested indexed map exposes two originals but remains outside
the flat-map splitter profile and is explicitly rejected. A partial REA graph
can demonstrate positive recovery; it cannot establish absence.
3. Inspect the representations real key consumers accept
The new key control executes Node key imports and byte-array decoding, plus
cryptography signing and authenticated decryption. The 19 inputs are specific
representations, not 19 independent vulnerability classes. They cover:
- JWK Base64 alphabet, folding, ignored characters, padding, unused bits and
Node's UTF-16 consumer aliases, including non-BMP and surrogate inputs. - Leading-zero RSA integer representations that preserve the actual key.
- Configured asset keys in integer-valued JSON numbers, contiguous windows in
longer byte arrays, folded Base64 and mixed-case hex. - Whole JSON strings containing supported unencrypted private DER in hex or
Base64, including ignored separators accepted by the decoding consumer.
Consumer recognition is distinct from canonical JWK validation. Supported JWK
families with nonempty declared private fields that cannot be completely
inspected now produce an explicit incomplete-inspection finding. They are not
counted as positively identified keys. Normal public keys remain allowed.
A separate 38,744-character ignored-byte representation still decodes to the
configured 32-byte key in Node. The known-key reader now checks a bounded raw
prefix: enough significant characters exclude an exact 32-byte whole-string
key, while an undecidable over-budget prefix explicitly blocks release. A
complete key followed by padding and a long ignored tail is still detected.
This regression is recorded as one budget rejection, outside the 19 positive
identifications.
JSON keeps its 8 MiB file, node and depth limits. Raw JWK fields are bounded
before allocating consumer normalization; whole-string known-key Base64 has a
32,768-character raw prefix budget. Encoded DER has a raw text inspection
bound, a 64 KiB per-candidate decoded limit, at most 64 candidates and a shared
256 KiB decoded-candidate budget per JSON document. Only private ASN.1 candidates
reach the cryptographic reader. Unsupported or over-budget recognized material
must not silently become a complete, successful scan. No secret values or JSON
property names are copied into findings.
Upgrade notes
The CLI/import names, SDK methods and signed-document v1 formats are unchanged.
Default release decisions become stricter in these cases:
- Finish final linking for Wasm files containing
linkingorreloc.*, then
separate debug information and test the application. Renamed files and
inspected ZIP/ASAR members receive the same metadata decision. - Remove private source content even from malformed or nonstandard JSON maps.
Merely naming the reserved content key may trigger the conservative rule. - Keep private key material and known external asset keys outside releases in
every accepted consumer representation. Resolve incomplete inspection; do not
treat it as a successfully checked public key. - Install matching 1.2.0 core/SDK archives and regenerate exported build assets
when upgrading. Sign and package only after all applicable checks succeed.
Required validation and downloadable evidence
The final local candidate ran 913 Python tests: 900 passed, 13 skipped, zero
failures. All 65 SDK tests and 10 real CLI integration tests passed, as did
fresh wheel, source and SDK installations. Local tool-dependent skips are
explicit; these results do not replace required CI. The local record
binds tested source hashes and identifies its uncommitted candidate state.
The required recovery CI job now runs both expanded corpora against npm/source,
the v1.1 key-representation control, and the existing v1.0 key, single-fixture
REA and Craft controls. Portable tests run on all three operating systems.
Every fresh wheel and source installation must also reject a consumer-compatible
key array, an escaped incomplete map and relocatable Wasm. The publisher requires
all three new installation booleans to be true.
Publication still requires all 13 named CI jobs at the exact main commit,
clean installed-artifact evidence, matching versions and verified uploaded
archive bytes. A skipped, missing, duplicate or unsuccessful required job blocks
publication. The release attaches installation-report.json,
release-evidence.json and SHA256SUMS alongside wheel, source and SDK archives.
Reproducible records
- Wasm corpus: npm and
pinned source. - JavaScript corpus: npm and
[pinned source](../results/sourc...
Anti-REA Forge v1.1.0
Anti-REA Forge 1.1.0 — Verified REA disclosure boundaries
Released 2026-11-18. Python core and JavaScript SDK both use version 1.1.0.
This release closes demonstrated information-disclosure paths with real REA
controls, supported private-metadata separation and bounded key inspection.
It does not claim that distributed algorithms cannot be reconstructed.
Measured changes
| Control | Before / original | After / public | Restoration or compatibility control |
|---|---|---|---|
| REA + WABT private Wasm identifiers | 3 recovered | 0 recovered | Restored private original: all 3 recovered |
| Wasm executable content | 255-byte original | 134-byte public module | Every standard section byte unchanged; 9 numerical results identical |
| REA source-map original text | 1 exact original recovered | 0 exact originals recovered | Private map restored: 1 exact original recovered; JS byte identity retained |
| Five recoverable key representations | v1.0 allows all 5; recovered keys decrypt or sign | v1.1 positively detects and rejects all 5 | Public JWK and ciphertext without its external key both remain accepted |
Both Wasm and source-map controls run against published REA 6.9.0 and an
exact clean source build of c338e54c33213ca8dd157d30fbc922b65a04c49b. These have
separate runtime identities and reports. Official WABT 1.0.42 asset and executable
hashes are checked. The source build's commit, tree, lock file, build tools and
complete dist inventory are recorded and checked before and after execution.
Errors, missing tools and incomplete analysis fail the control.
Three private identifiers belong to one owned module; this is not three
independent attacks or a general success-rate estimate. Readable arithmetic WAT
remains present in every group. The five key controls use standard decoders and
cryptography rather than an REA agent session.
What changed
WebAssembly metadata and private originals
Default portable inspection recognizes bounded core-v1 Wasm envelopes by magic,
including renamed files and inspected ZIP/ASAR members. It rejects supported
private name, .debug_*, .zdebug_*, sourceMappingURL and
external_debug_info custom sections. Malformed envelopes, unsupported versions,
ambiguous private sections and exceeded budgets receive an explicit failure.
The new command:
mkdir public-wasm private-wasm
python -m releaseguard split-wasm-debug \
--input build/pricing.wasm \
--public public-wasm/pricing.wasm \
--private private-wasm/pricing.original.wasmcreates two new files. The private file is the complete original module;
only the public file belongs in the release. All retained sections are compared
byte-for-byte, both outputs are read back, and failed creation or changed input
rolls back newly created files. Unknown custom sections are preserved. The
production scanner/splitter does not require Node or WABT and never executes the
input. Complete integration and limits
Decoded JSON and DER key disclosure
The same bounded JSON read now checks source maps and supported key material.
It inspects escaped strings, root-string documents, object property names, exact
configured 32-byte arrays and supported JWK private components. Public fields
cannot hide directly recoverable private components merely by being absent or
incorrect. Unknown expressions, split/derived keys and arbitrary encodings remain
outside this scope; no target JavaScript is evaluated.
Standalone unencrypted DER recognition first reads a small ASN.1 prefix and only
admits private-key-shaped candidates to a bounded decoder. Supported private keys
are rejected. Recognized oversized or uninspectable private structures fail
explicitly. Public-key DER, public JWK and certificates have compatibility controls.
Report fields never include secret values or decoded JSON property names.
Installation and publication
Fresh wheel and source installations must perform Wasm separation and decoded-key
rejection on all three operating systems. Real REA npm and pinned-source recovery
controls are mandatory in the existing 13-job release policy. Publication still
requires a successful exact main commit, clean installation evidence, exact asset
hashes and a verified draft upload before making the Release public.
Upgrade
- CLI/import names and existing v1 signed-document and policy formats remain the same.
check_native_metadata: truenow rejects recognized Wasm private metadata. Run the new splitter or change your build to stop emitting private sections, then test the actual application.check_private_keys: truenow checks the supported decoded JSON/JWK and DER representations. Supplyingasset_key_filesalso enables their decoded JSON checks even when generic private-key marker checks are disabled.- Reports add
structured_secrets_checked,der_private_keys_checkedandstructured_secret_scan_completed. Any positive or incomplete-recognition finding keeps the gate closed. - Keep the original module, private map, credentials and issuer materials outside all public archives and deployments. Splitting must happen before signing or hashing the released module.
Evidence and limits
- npm Wasm control
- pinned-source Wasm control
- npm source-map control
- pinned-source source-map control
- Five key recoveries and two nonsecret controls
- Capability review and explicit remaining gaps (中文)
Wasm framing checks do not validate all instruction/type semantics. Custom-section
introspection can change application behavior, so developers must test their
actual embedding. PYC, CIL and public JS/Wasm instructions remain analyzable;
authorized runtime memory remains observable. The trusted service/external-key
boundary is still necessary for secrets that must not be delivered to a client.
See SECURITY.md for the full threat model. Downloaded assets
include release-evidence.json, installation-report.json and SHA256SUMS, which
bind the published files to the final main-branch CI run.
Anti-REA Forge v1.0.0
Anti-REA Forge v1.0.0
2026-11-18 · English README · 中文 README
Release focus
Version 1.0 combines a broader release inspector, verified native metadata separation,
stronger build/admission boundaries and an installation-tested distribution pipeline.
Python and JavaScript package versions are aligned at 1.0.0. Existing v1 signed
formats, the releaseguard CLI/import name and SDK entry points are retained.
LLVM native function transformation remains experimental; this release is not an
independent security audit or a claim that general-purpose analysis tools cannot run.
Audit findings and reproductions
The audit began at 674f8b20cc0fae7bd2684b15dc74cc4db567db29. Each row represents
a concrete reproduced behavior with owned inputs, followed by a production change
and both rejection and compatibility controls.
| Area | Before | 1.0 behavior | Reproduction / regression |
|---|---|---|---|
| ELF CTF/BTF | GCC type metadata survived GNU strip; native and release checks reported no debug disclosure | Exact type/index sections detected; eligible non-allocated bytes retained privately and removed publicly | Real GCC/readelf/objcopy controls |
| Source-map encodings and wrappers | UTF-16/32 XSSI maps and root-array embedded maps could pass inspection | Encoded delimiters aligned by code unit; objects and arrays inspected under existing byte/node/depth budgets | Encoding and JSON controls |
| Opaque packages | Renamed canonical V7 TAR and expected Brotli sidecars were treated as ordinary bytes | V7 structure/checksum recognized and .br suffix rejected as unsupported | Independent TAR/Node Brotli round trips |
| CMake gate identity | cwd/PYTHONPATH shadows replaced native-init and made a compiled forbidden-literal program pass POST_BUILD | Selected installed interpreter uses isolated mode for both commands | Four actual CMake shadow controls |
| Final LLVM emission closure | Direct sealing input retained prefix/prologue constants and inline/module assembly outside inventoried storage | Actual syntax tokens rejected across the module while identifiers, strings and comments remain valid | 11 emission variants and real LLVM controls |
| Native static teardown | A late global destructor could reopen the generated loader after its owner had been destroyed | Automatic teardown retires the instance; normal explicit clear/reinitialize still works | Actual GCC/OpenSSL lifetime controls |
| Capability admission timing | A capability could expire during durable commit/close yet reach the handler | Trusted time sampled again after consumption; final denial leaves the nonce spent | Real Ed25519/SQLite completion controls |
| Replay-store identity | One live store object accepted a fresh or byte-identical replacement database | Instance pins device/inode and rejects ordinary replacement | Replacement and normal multi-instance controls |
| Bounded document input | Legacy reads could block on FIFO or read without a bound before rejecting a negative budget | Regular, stable files only; bounded reads and pre/post identity checks | FIFO, links, changes, budgets and ordinary documents |
The LLVM emission issue is post-transform/direct-API hardening: original-source
preflight already rejected those constructs, so this is not presented as a
proven ordinary CLI bypass. Replay identity checks do not provide rollback-proof
storage. TAR/MSF/ar/Brotli recognition is explicit incomplete-inspection handling,
not a new parser or proof that the contents are protected. Arbitrarily renamed
Brotli streams have no supported magic recognition.
Release acceptance and downloadable evidence
CI has 13 named required jobs across Linux,
Windows and macOS. The publisher requires an
entire successful main push run at the current main SHA, with the exact job
inventory in release-policy.json. PR/fork runs,
skipped jobs, changed commits and incomplete job lists cannot publish.
Three packaging jobs build and install the actual wheel and Python source archive
in separate fresh environments outside the checkout. They verify isolated imports,
CLI version, shipped native resources, clean/blocked releases, a 4 KiB authenticated
asset round trip and leaked-key rejection. The packed SDK is installed in a separate
Node project and exercised against the installed Python core.
Publication uses the exact Linux job archives. The other operating systems independently build and install the same source; they do not install identical Linux-produced archive bytes. Their names, byte sizes and SHA-256
are checked against the installation report. A new tag identifies the validated
commit; assets are uploaded into a draft, and server-reported asset digests are
checked before making the release public. Existing releases or tags are never
replaced. A failed upload/check remains unpublished for review.
The v1.0.0 Release
contains:
open_release_guard-1.0.0-py3-none-any.whlopen_release_guard-1.0.0.tar.gzanti-rea-forge-releaseguard-1.0.0.tgzinstallation-report.json— exact package checks and archive hashesrelease-evidence.json— final commit, all required CI job URLs and installation reportSHA256SUMS— checksums for all five files above
The Python and SDK archives are GitHub Release assets. This publication does not
claim a PyPI or npm registry release. CI success is operational evidence, not a
cryptographic signature or independent security certification. The release record
is generated from actual final-main results, rather than copied from an earlier PR.
Upgrade from 0.10 / SDK 0.6
- Install the core into the selected interpreter. CMake now uses
python -I;
PYTHONPATH and user-site-only installations are excluded. Use a virtual
environment, install the package there and setPython3_EXECUTABLEfor CMake.
Regenerate previously exported CMake assets so builds use the new gate. - Update the dependency baseline. The core requires
cryptography>=50.0.2,<51.
Upstream fixed security issues and updated bundled OpenSSL after the former
allowed floor. This is dependency hygiene; no reachable PKCS7/X509 exploit in
ReleaseGuard was demonstrated. Python 3.10+ remains supported. Intel macOS and
32-bit Windows no longer receive wheels from recent cryptography releases;
our macOS/Windows runner matrix does not establish support for those targets. - Use regular files for documents, policies and PEM keys. Unsafe leaf links,
reparse points, FIFOs and other special files are rejected. Parent directories
must remain trusted. The internal bounded reader accepts a plain integer in
0..8 MiB; individual callers retain their smaller limits. - Handle additional release findings. CTF/BTF sections now fail the default
native metadata check. Usesplit-elf-debugon eligible original linked ELF
input, retaining its private sidecar..br, recognized V7 TAR and previously
unsupported formats must be removed, inspected in an appropriate separate
workflow or covered by an explicit reviewed policy decision. An opt-out is
not equivalent to having inspected their contents. - Keep authorization retries explicit. A token rejected by the final time
check remains spent. Obtain a new token rather than resetting replay storage.
The database schema and existing signed capability format are unchanged. - Respect native lifetime order. Initialize before workers, clear after
joining them. Automatic module destruction is final; late destructors cannot
reopen it. Public runtime/helper code and authorized plaintext remain observable.
Sources for the dependency change: cryptography 50.0.2 changelog
and published package metadata.
What the data does and does not show
The README links the historical paired Ghidra/IDA asset experiment, with plaintext,
external-key and deliberately leaked-key controls, candidate counts and patch results.
Its 1 MiB latency number is a dated single-host observation, not a universal benchmark.
Current CI also runs pinned REA CLI/source-map recovery, genuine Ghidra extraction,
actual LLVM/native compilers and official ASAR controls. Missing local toolchains
are recorded as skips and never counted as execution.
There is no new claim of modern IDA/Hex-Rays, Hopper, Frida, GDB, JADX, JEB, firmware
or complete REA-agent/model resistance. The controls reduce shipped disclosures and
close specific admission/build/format errors. They cannot revoke an independent
reader's access to already distributed code or prevent authorized runtime observation.
中文升级摘要
本轮修复 9 类已复现的行为缺口,保留正常使用的正对照;Python 与 SDK 统一为 1.0.0。
新的发布门禁要求 13 个指定 CI 作业成功,并核对实际安装包及上传文件的 SHA-256。
中英文 README 的原理图由仓库内 SVG 生成脚本
维护,私有密钥、公开产物和授权执行的边界在图中分别标明。
升级时需要将内核安装进 CMake 选定的解释器,重新生成旧版 CMake 资产;普通文档、策略、
PEM 必须为可信普通文件。CTF/BTF、.br、V7 TAR 等新增检查可能使旧发布流程失败,需处理
真实泄露或明确评审策略。授权提交后过期会拒绝进入处理函数,令牌保持已消费;重放数据库
仍依赖可信持久化存储。原生 LLVM 转换继续标为实验功能,不宣称所有通用逆向工具都无法突破。