Repository navigation
Releases
v0.11.0
Added
- Test-verified fixes.
nephilex fix --verify-with-tests(with
--test-command, orautofix.test_commandfrom a trusted config) runs the
project's tests in a temporary git worktree before and after the fixes,
re-scans, and commits only fixes that keep the tests passing. The command
runs without a shell, with a timeout and capped output, and with tokens and
keys removed from its environment. Results appear in the summary, the commit,
the pull request body and the audit log. - Agent-security rules.
llm.indirect-prompt-injection(external content
in the prompt of a model whose tools or output can act on it),
llm.tool-argument-to-sink,llm.untrusted-message-role,
llm.untrusted-tool-selectionandllm.secret-sent-to-model, with
JavaScript/TypeScript coverage for untrusted system prompts and tool handlers. - Benchmarks with held-out data.
benchmarks/heldout/scores Nephilex,
Bandit and Semgrep on 201 files from real GitHub advisory fix commits,
before and after each fix;benchmarks/agent/scores the LLM and agent
rules on a labeled corpus with a held-out split. Both publish raw results. - Documentation: a quickstart, one finding followed from detection to fix, and
a "Trust and limits" page with a capability matrix. Scanner.unavailable_reason()lets a scanner say why it is skipped.
Changed
- The GitHub Action installs
nephilex[ast], so the taint scanners run in CI.
Earlier versions installed it without the extra and skipped them. - A scan without the
[ast]extra says the taint scanners need it, instead of
calling them "not applicable". - When two tiers report one weakness with equal confidence and severity, the
finding with a deterministic fix is kept, then the AST one. The choice no
longer depends on scanner order, soscanandcheckreport the same rule
andcheckkeeps the fix an editor applies. - The weak-hash fix is not applied to a line that hashes a password or other
secret: SHA-256 is still a fast hash, and passwords need a key-derivation
function. Such findings keep that guidance instead of a "verified" patch.
Fixed
--formathelp onscanandimportnow listsjunit.- Documentation that did not match the code: the number of report formats,
--live-targetdescribed as checking TLS (it checks HTTPS and the redirect),
and dependency reachability and VEXnot_affected, which need
--reachability. - Deeply nested YAML in
.nephilex.yml,.nephilex/context.yml,
.nephilex/exceptions.yml, custom rule files, a scanned project's own config
(read by the DAST scanner) or the agent config no longer crashes the run with
aRecursionError. Configuration files fail with their normal error, and
rule or project files are skipped. Found by the new fuzz tests.
Security
- CI, the release workflow, the container image and the fuzzing build install
every dependency from hash-locked files (requirements/), and Nephilex is
built without fetching an unpinned build backend. - Each release carries its signed build provenance (
nephilex-<version>.intoto.jsonl)
next to the SBOM, verifiable with
gh attestation verify <file> --bundle nephilex-<version>.intoto.jsonl --owner Nephilex. - Coverage-guided fuzzing (ClusterFuzzLite with atheris) of the SARIF,
config and context parsers runs on pull requests and weekly.
v0.10.0
The first release under the Nephilex name. It contains the 0.9.0 security
hardening, the 0.9.2 fixes, which were never published, and the changes below.
Breaking changes
-
The project is now Nephilex. Everything that carries the product name
changed with it, with no aliases for the earlier names:- the PyPI package and the command are
nephilex, and the import package
isnephilex(plugins importnephilex.plugin_apiand register under the
nephilex.commandsentry-point group); - repository config is
.nephilex.ymland.nephilex/(rules, exceptions,
context, audit log), and the inline suppression comment is
nephilex-ignore; - environment variables use the
NEPHILEX_prefix (for example
NEPHILEX_CACHE_DIR,NEPHILEX_MCP_ROOTS,NEPHILEX_DASHBOARD_TOKEN); - the MCP server is registered as
nephilex, and its tools are
nephilex_scan,nephilex_check_filesand so on; - SARIF output carries the fingerprint as
nephilexFingerprint, so baselines
and code-scanning alerts made by earlier versions need to be regenerated; - the GitHub repository is
nephilex/nephilex, the Action is
nephilex/nephilex@<tag>, and the container image is
ghcr.io/nephilex/nephilex.
To upgrade, install
nephilex, rename the config file and directory,
update inline suppression comments, CI workflows and environment variables,
and rerunnephilex init --editorfor editor integrations. - the PyPI package and the command are
Added
- SARIF import.
nephilex import <file.sarif>andnephilex scan --import-sarif
read SARIF 2.1.0 from other scanners (CodeQL, Semgrep, Bandit and others),
map their results onto Nephilex findings, and run them through the same
baseline, suppression, policy and reporting steps as native findings. The
input is treated as untrusted: file size, nesting and result count are
capped, locations outside the project root are dropped, and snippets are
redacted before anything is stored. - Project context file.
.nephilex/context.ymldeclares the project's own
sanitizers and trusted sources (used by the taint engines), internal-only
paths, exposure levels and data classification, so findings are ranked
against how the code is really deployed. Validation is strict, and like the
rest of repository config it is read only from trusted projects. - Audit log.
.nephilex/audit.logis an append-only, hash-chained record of
scans, suppressions (including findings hidden by--baseline) and fixes.
Setaudit_log: falseto turn it off. - LLM and agent data flow. The LLM scanner parses Python with tree-sitter
to trace model output into dangerous sinks (LLM05), flag agents given
excessive tools (LLM06) and review system-prompt construction. Findings cite
OWASP Top 10 for LLM Applications 2025 ids. - Agent configuration scanner (
agent-config). Reviews MCP server
definitions and coding-agent settings for hardcoded credentials, servers
launched from unpinned packages, remote servers over plain HTTP, filesystem
access to/or a home directory, and settings that run tools without
confirmation. It runs in thedeepprofile. nephilex check. A fast, file-level check for editors and AI assistant
hooks, with--hookoutput for Claude Code and Cursor and
--no-inline-ignoresfor strict mode.- MCP tools. The MCP server adds
nephilex_check_files,nephilex_check_snippet,
nephilex_check_packageandnephilex_fixnext tonephilex_scan. nephilex init --editor claude|cursor|vscodewrites that editor's MCP and
hook configuration into the project.- Public benchmarks.
benchmarks/public/runs Nephilex, Bandit and Semgrep
on OWASP BenchmarkPython and SecurityEval and publishes the raw findings and
scores; seedocs/benchmarks.md. - Flow-sensitive Python taint engine. Branches, loops,
matchandtry
are analysed separately and joined; constant conditions prune dead code;
literal containers and simple early-return guards are understood. New
CWE-mapped sinks cover XSS, open redirect, XXE, LDAP and XPath injection,
trust-boundary violations, weak randomness for tokens, unsafe
deserialization and insecure cookies. - Findings report start and end columns, and snippets keep their indentation,
so editors can highlight the exact match. - Logo set and brand assets in
docs/assets/brand/.
Changed
- A scan of a git work tree skips untracked files that git ignores (build
output, virtual environments, local.envfiles). - Inline
nephilex-ignorecomments are recognised in--,<!-- -->and;
comment styles as well as#and//. - Deterministic fix patches are attached when AI is off, including by
nephilex check, so editors can offer the verified fix in place. - Reporters, model clients and HTTP libraries are imported on first use, so
the CLI starts faster. - The regex tier asks the taint engine about its payload, so
shell=True,
os.systemandpickle.loadson literal-only data are no longer reported.
Fixed
- Imported SARIF findings keep the tool's own fingerprints, so baselines stay
stable across runs. nephilex mcp --helplists every MCP tool.- A fresh
import nephilex.inventoryno longer fails on a circular import.
v0.9.1
Do not use this release. The v0.9.1 tag was created on the 0.9.0 commit by mistake. The PyPI release check refused to publish it, so it is not on PyPI, and its container image contains 0.9.0. Use 0.9.2.
Changelog
All notable changes to Argus are documented here. The format follows
Keep a Changelog and the project adheres
to Semantic Versioning (pre-1.0: minor versions may
include breaking changes, noted explicitly).
[Unreleased]
[0.9.1] - 2026-10-22
0.9.0 was tagged and its container image published, but the PyPI upload was
rejected (see below), so 0.9.1 is the first PyPI release with the 0.9.0
security hardening. It contains every 0.9.0 change plus these fixes.
Fixed
- Packages declare core metadata 2.4 instead of 2.5, and releases upload with
pypa/gh-action-pypi-publishv1.14.3. The upload tooling pinned for 0.9.0
could not read metadata 2.5 and rejected the distributions. - The curl-pipe-shell rule no longer fires on text that only names the
pattern (for example "avoidcurl | bash" in documentation or a rule's own
explanation); a real download-and-execute always has an argument between
the command and the pipe. - The scan cache now includes a fingerprint of the installed Argus code, so
changing a rule in a source checkout no longer returns stale cached findings.
[0.9.0] - 2026-10-22
This release is a security hardening release. Argus runs on untrusted code (pull
requests from forks, cloned repositories), and a full review found ways a
hostile repository could execute code, read outside the scan root, hang a
scan, hide its own findings, or make a CI gate pass. All of them are fixed.
Upgrading is recommended for every user.
Breaking changes
- Repository config is untrusted by default.
Config()now defaults to
trust_project_config=False. A cloned remote's.argus.yml,
.argus/rules,.argus/exceptions.ymland inlineargus-ignorecomments
are ignored unless you pass--trust-remote-config. Library code that builds
Config()directly must settrust_project_config=Trueto honour them. - Scanner crashes fail the gate. When a
fail_onthreshold is set, a
scanner error now fails the run (a crashed scanner reports nothing, which a
gate must not read as a pass). Opt out with--no-fail-on-erroror
fail_on_error: false.policy checkalways fails on scanner errors. - Configuration mistakes exit 2. Wrong types in
.argus.yml(for example
exclude_paths: "dist/*"), unknown--scannersnames, a selection that
leaves no scanner, an unknown profile, an invalid severity, invalid policies,
a malformedexceptions.yml, andpolicy checkwith no policies configured
now exit 2 instead of being coerced or ignored. - Dashboard requests are restricted. Only loopback Host headers are
accepted (add names withARGUS_DASHBOARD_ALLOWED_HOSTS);/api/scans
requiresContent-Type: application/json; the upload form requires its CSRF
token; binding to a non-loopback address requiresARGUS_DASHBOARD_TOKEN. argus pushrefuseshttp://except for loopback hosts.- The MCP server only scans inside its working directory (or
ARGUS_MCP_ROOTS). - A DAST URL in the scanned project's own
.argus.ymlis ignored. Use
--live-targetorARGUS_DAST_URL. argus fixruns your git commit hooks and commits with your own git
identity. It refuses to run whenHEADis not the base branch.watchstate directory names changed, so the first cycle after
upgrading records a fresh baseline.
Security
- Scanning a repository can no longer execute code through its
.git/config
(core.fsmonitor,gpg.program,diff.external, textconv); every git call
runs with hardened config overrides, timeouts and bounded output. - GitHub Action inputs are passed to the shell through the environment instead
of being interpolated, closing a command-injection path for workflows that
forward untrusted event data. On pull requests the Action now reads Argus
config from the base branch (pr-config-from-base). - Regular expressions that a crafted file could drive into minutes of
backtracking were rewritten to linear forms, and every line-based search is
bounded. - Secret values are masked everywhere they appear in a snippet, including
second tokens on the same line and CI/CD and container findings. - Dashboard: DNS-rebinding and cross-site request protection, security headers,
optional token authentication, request size limits, owner-only database
permissions, and clamped report timestamps. - DAST and live posture probes verify the connected peer address, refuse
redirects into private networks, and cap response size. - Markdown reports, PR review comments and fix PR bodies escape untrusted text,
links, HTML and @mentions. - Every prompt sent to a model fences and caps the scanned snippet;
AI-proposed patches are never marked verified. - Remote clones are bounded, skip Git LFS downloads, and never leak temporary
directories on timeout.
Added
--config-from-ref REFonscan,review,policy checkand
baseline create: read Argus config from a git ref (the base branch) and
treat the checked-out tree as untrusted.fix --no-verifyto skip commit hooks explicitly.ai.max_findings(default 25): only the highest-risk findings are sent to a
model; the rest get offline reasoning.ARGUS_AI_TIMEOUT,ARGUS_AI_MAX_RETRIES,ARGUS_GIT_TIMEOUT,
ARGUS_MCP_SCAN_TIMEOUT,ARGUS_DASHBOARD_MAX_SCANS,
ARGUS_WATCH_KEEP_SNAPSHOTSandARGUS_DAST_URL.- New informational findings when part of an analysis could not run:
<scanner>.analysis-skipped,secrets.history-scan-failed,
secrets.history-scan-incomplete,agent.step-failed,fim.file-created,
fim.unreadable,host.ssh-config-unreadable,host.check-failed. - Python 3.13 support, CodeQL and dependency review in CI, signed build
provenance and SBOM attestations for releases and the container image. - Governance:
GOVERNANCE.md,MAINTAINERS.md,SUPPORT.md, a complete
security policy, the full Contributor Covenant, and DCO sign-off. - Versioned plugin API:
argus.plugin_apiis the public import surface for
scanners, reporters and AI providers, versioned byPLUGIN_API_VERSION(1).
Plugin base classes carryapi_version; a plugin built for a different
version is skipped with a logged warning instead of being loaded. See
"Stability and versioning" indocs/plugins.md.
Changed
ScannerContextis now a dataclass. Construction and attribute access are
unchanged.- The CLI is split into one module per command group under
argus.cli.commands; commands, options, help text and exit codes are
unchanged.argus.cli.main:appremains the entry point. - All git calls go through one hardened runner,
argus.core.gitexec. The fix
and review workflow, the dependency diff and--config-from-refnow get the
same config pins, timeouts and output caps as the git-based scanners. Git
never prompts for credentials (GIT_TERMINAL_PROMPT=0). - The default Anthropic model is
claude-sonnet-5-5(override with
ARGUS_AI_MODEL,ai.modelor--ai-model); model calls default to a 60 s
timeout and 2 retries; the defaultmax_tokensis 4096. argus watchre-reads the target every cycle, keeps running after a failed
cycle, and prunes old snapshots. The host agent keeps monitoring when a step
or push fails, and reports the failure.- The
ciprofile disables the scan cache. --diffkeeps every finding (with a warning) when the diff cannot be
computed, and works when scanning a subdirectory.- PR review comments use repository-relative paths and the summary is capped
at 50 entries. - The malware scanner reports findings in test files at lower severity
instead of skipping them, and never skipsconftest.pyorsetup.py. - The container image is pinned by digest and installed from the hash-locked
dependency set.
Fixed
sbomandinventoryJSON on stdout is no longer wrapped or stripped.argus watchsaw only the first cycle's files.- The host agent never detected listening ports on Linux and crashed on
unreadablesshd_config; file integrity monitoring missed new files,
directory contents and unreadable files. - Packages beyond a fixed count were silently dropped from OSV lookups.
- Deeply nested files or malformed manifests could zero out a whole scanner or
crash the scan. - The auth map ranked risks in the wrong order; PostgreSQL was detected from
"jpg" and "upgrade"; the reachability risk boost and the AWS credential
attack chain never fired. - JUnit and SARIF output stay valid for any path or text; GitLab report ids are
unique. - The fix applier keeps original line endings, skips non-UTF-8 files, and
restores the working tree and branch on failure. argus dependencies,argus supply-chainandargus sbomexit 2 with a
message for an unknown or option-like--diffref instead of a traceback.ScanEngine.scanno longer modifies theProjectpassed in, so its
exclude_pathsno longer carry over to a later scan of the same object.
[0.8.1] - 2026-09-07
Added
malwarescanner: backdoors, webshells, obfuscated exec, reverse shells,
cryptominers, credential exfiltration, suspicious npm install scripts.lang-sast: Go, C#, Kotlin, Swift, shell, and PowerShell rules.
Fixed
- Deep profile:
ast-python-interprocrenamed toast-python-xfile(cross-file taint). - Authz precision: skip test files and public auth routes; detect router-level
Depends(); reduce false positives on token-in-URL heuristics.
[0.8.0] - 2026-09-07
Platform expansion: su...
v0.9.0
Note: this version was not published to PyPI. The upload was rejected because of a packaging metadata incompatibility. Install 0.9.1 or later, which contains every change below.
Changelog
All notable changes to Argus are documented here. The format follows
Keep a Changelog and the project adheres
to Semantic Versioning (pre-1.0: minor versions may
include breaking changes, noted explicitly).
[Unreleased]
Fixed
- The curl-pipe-shell rule no longer fires on text that only names the
pattern (for example "avoidcurl | bash" in documentation or a rule's own
explanation); a real download-and-execute always has an argument between
the command and the pipe. - The scan cache now includes a fingerprint of the installed Argus code, so
changing a rule in a source checkout no longer returns stale cached findings.
[0.9.0] - 2026-10-22
This release is a security hardening release. Argus runs on untrusted code (pull
requests from forks, cloned repositories), and a full review found ways a
hostile repository could execute code, read outside the scan root, hang a
scan, hide its own findings, or make a CI gate pass. All of them are fixed.
Upgrading is recommended for every user.
Breaking changes
- Repository config is untrusted by default.
Config()now defaults to
trust_project_config=False. A cloned remote's.argus.yml,
.argus/rules,.argus/exceptions.ymland inlineargus-ignorecomments
are ignored unless you pass--trust-remote-config. Library code that builds
Config()directly must settrust_project_config=Trueto honour them. - Scanner crashes fail the gate. When a
fail_onthreshold is set, a
scanner error now fails the run (a crashed scanner reports nothing, which a
gate must not read as a pass). Opt out with--no-fail-on-erroror
fail_on_error: false.policy checkalways fails on scanner errors. - Configuration mistakes exit 2. Wrong types in
.argus.yml(for example
exclude_paths: "dist/*"), unknown--scannersnames, a selection that
leaves no scanner, an unknown profile, an invalid severity, invalid policies,
a malformedexceptions.yml, andpolicy checkwith no policies configured
now exit 2 instead of being coerced or ignored. - Dashboard requests are restricted. Only loopback Host headers are
accepted (add names withARGUS_DASHBOARD_ALLOWED_HOSTS);/api/scans
requiresContent-Type: application/json; the upload form requires its CSRF
token; binding to a non-loopback address requiresARGUS_DASHBOARD_TOKEN. argus pushrefuseshttp://except for loopback hosts.- The MCP server only scans inside its working directory (or
ARGUS_MCP_ROOTS). - A DAST URL in the scanned project's own
.argus.ymlis ignored. Use
--live-targetorARGUS_DAST_URL. argus fixruns your git commit hooks and commits with your own git
identity. It refuses to run whenHEADis not the base branch.watchstate directory names changed, so the first cycle after
upgrading records a fresh baseline.
Security
- Scanning a repository can no longer execute code through its
.git/config
(core.fsmonitor,gpg.program,diff.external, textconv); every git call
runs with hardened config overrides, timeouts and bounded output. - GitHub Action inputs are passed to the shell through the environment instead
of being interpolated, closing a command-injection path for workflows that
forward untrusted event data. On pull requests the Action now reads Argus
config from the base branch (pr-config-from-base). - Regular expressions that a crafted file could drive into minutes of
backtracking were rewritten to linear forms, and every line-based search is
bounded. - Secret values are masked everywhere they appear in a snippet, including
second tokens on the same line and CI/CD and container findings. - Dashboard: DNS-rebinding and cross-site request protection, security headers,
optional token authentication, request size limits, owner-only database
permissions, and clamped report timestamps. - DAST and live posture probes verify the connected peer address, refuse
redirects into private networks, and cap response size. - Markdown reports, PR review comments and fix PR bodies escape untrusted text,
links, HTML and @mentions. - Every prompt sent to a model fences and caps the scanned snippet;
AI-proposed patches are never marked verified. - Remote clones are bounded, skip Git LFS downloads, and never leak temporary
directories on timeout.
Added
--config-from-ref REFonscan,review,policy checkand
baseline create: read Argus config from a git ref (the base branch) and
treat the checked-out tree as untrusted.fix --no-verifyto skip commit hooks explicitly.ai.max_findings(default 25): only the highest-risk findings are sent to a
model; the rest get offline reasoning.ARGUS_AI_TIMEOUT,ARGUS_AI_MAX_RETRIES,ARGUS_GIT_TIMEOUT,
ARGUS_MCP_SCAN_TIMEOUT,ARGUS_DASHBOARD_MAX_SCANS,
ARGUS_WATCH_KEEP_SNAPSHOTSandARGUS_DAST_URL.- New informational findings when part of an analysis could not run:
<scanner>.analysis-skipped,secrets.history-scan-failed,
secrets.history-scan-incomplete,agent.step-failed,fim.file-created,
fim.unreadable,host.ssh-config-unreadable,host.check-failed. - Python 3.13 support, CodeQL and dependency review in CI, signed build
provenance and SBOM attestations for releases and the container image. - Governance:
GOVERNANCE.md,MAINTAINERS.md,SUPPORT.md, a complete
security policy, the full Contributor Covenant, and DCO sign-off. - Versioned plugin API:
argus.plugin_apiis the public import surface for
scanners, reporters and AI providers, versioned byPLUGIN_API_VERSION(1).
Plugin base classes carryapi_version; a plugin built for a different
version is skipped with a logged warning instead of being loaded. See
"Stability and versioning" indocs/plugins.md.
Changed
ScannerContextis now a dataclass. Construction and attribute access are
unchanged.- The CLI is split into one module per command group under
argus.cli.commands; commands, options, help text and exit codes are
unchanged.argus.cli.main:appremains the entry point. - All git calls go through one hardened runner,
argus.core.gitexec. The fix
and review workflow, the dependency diff and--config-from-refnow get the
same config pins, timeouts and output caps as the git-based scanners. Git
never prompts for credentials (GIT_TERMINAL_PROMPT=0). - The default Anthropic model is
claude-sonnet-5-5(override with
ARGUS_AI_MODEL,ai.modelor--ai-model); model calls default to a 60 s
timeout and 2 retries; the defaultmax_tokensis 4096. argus watchre-reads the target every cycle, keeps running after a failed
cycle, and prunes old snapshots. The host agent keeps monitoring when a step
or push fails, and reports the failure.- The
ciprofile disables the scan cache. --diffkeeps every finding (with a warning) when the diff cannot be
computed, and works when scanning a subdirectory.- PR review comments use repository-relative paths and the summary is capped
at 50 entries. - The malware scanner reports findings in test files at lower severity
instead of skipping them, and never skipsconftest.pyorsetup.py. - The container image is pinned by digest and installed from the hash-locked
dependency set.
Fixed
sbomandinventoryJSON on stdout is no longer wrapped or stripped.argus watchsaw only the first cycle's files.- The host agent never detected listening ports on Linux and crashed on
unreadablesshd_config; file integrity monitoring missed new files,
directory contents and unreadable files. - Packages beyond a fixed count were silently dropped from OSV lookups.
- Deeply nested files or malformed manifests could zero out a whole scanner or
crash the scan. - The auth map ranked risks in the wrong order; PostgreSQL was detected from
"jpg" and "upgrade"; the reachability risk boost and the AWS credential
attack chain never fired. - JUnit and SARIF output stay valid for any path or text; GitLab report ids are
unique. - The fix applier keeps original line endings, skips non-UTF-8 files, and
restores the working tree and branch on failure. argus dependencies,argus supply-chainandargus sbomexit 2 with a
message for an unknown or option-like--diffref instead of a traceback.ScanEngine.scanno longer modifies theProjectpassed in, so its
exclude_pathsno longer carry over to a later scan of the same object.
[0.8.1] - 2026-09-07
Added
malwarescanner: backdoors, webshells, obfuscated exec, reverse shells,
cryptominers, credential exfiltration, suspicious npm install scripts.lang-sast: Go, C#, Kotlin, Swift, shell, and PowerShell rules.
Fixed
- Deep profile:
ast-python-interprocrenamed toast-python-xfile(cross-file taint). - Authz precision: skip test files and public auth routes; detect router-level
Depends(); reduce false positives on token-in-URL heuristics.
[0.8.0] - 2026-09-07
Platform expansion: supply chain, policy engine, SBOM, infrastructure scanning,
and the argus.commands entry point for add-on CLI commands.
Added
- Supply chain: malicious package intel, typosquat detection, npm behavior
metadata,argus supply-chaincommand. - SBOM: CycloneDX and SPDX output, SBOM diff via
argus sbom --diff. - Policy engine:
argus policy checkwith YAML policies. - Dependency diff:
argus dependencies --difffor pull-request review. - **Infrastructure sc...
v0.8.1
Release 0.8.1: SBOM push, ruff fixes, action pin updates
v0.8.0
Changelog
All notable changes to Argus are documented here. The format follows
Keep a Changelog and the project adheres
to Semantic Versioning (pre-1.0: minor versions may
include breaking changes, noted explicitly).
[Unreleased]
[0.8.0], 2026-09-07
Platform expansion: supply chain, policy engine, SBOM, infrastructure scanning,
and plugin CLI commands for commercial add-ons.
Added
Supply chain: malicious package intel, typosquat detection, npm behavior
metadata, argus supply-chain command.
SBOM: CycloneDX and SPDX output, SBOM diff via argus sbom --diff.
Policy engine: argus policy check with YAML policies.
Dependency diff: argus dependencies --diff for pull-request review.
Infrastructure scanning: cicd, container, and cloud scanners;
focused CLI commands argus cicd, argus container, argus cloud.
Additional scanners: authz, API security (OpenAPI), provenance, dependency-diff.
Reachability: npm import reachability and Python symbol-level tier.
Docs: repository split, project layout, infrastructure scanning, publishing checklist.
Changed
Version 0.8.0 is the minimum core release for argus-k8s plugin commands.
[0.7.0], 2026-07-14
The largest release so far: AI-era security coverage, deeper code analysis,
signal-quality features, and frictionless distribution. It also folds in the
JavaScript/TypeScript AST scanner and the web dashboard, which were built but
never released.
Added
AST data-flow (taint) scanner for JavaScript/TypeScript via tree-sitter:
follows untrusted input through multiple hops into sinks, and treats
parameterized queries and sanitized values as safe.
Web dashboard (optional extra [dashboard]): a FastAPI and SQLModel app
with scan history and risk trends, started with argus dashboard.
LLM / AI-application security scanner (llm), mapped to the OWASP Top 10
for LLM Apps: insecure handling of model output (taint pass into
eval/exec/shell/SQL/HTML sinks), prompt injection, secrets in prompts,
over-privileged agent tools, trust_remote_code=True, unsafe torch.load
pickle loading, and HTTP model downloads. Runs only on files that use an
LLM/agent stack. See docs/llm-security.md.
Secret verification (opt-in, --verify-secrets): read-only calls confirm
whether a detected credential is actually live (GitHub, Stripe, Slack,
OpenAI, Google), escalating confirmed ones to Critical and clearing format
false positives. Local targets only, never in CI templates; the secret is
never stored in a finding, cache, or report.
Live-target posture checks (--live-target , and argus scan ):
safe, read-only, non-intrusive runtime checks, security headers, cookie
flags, HTTP-vs-HTTPS transport, version disclosure, and a short allowlist of
exposed sensitive paths (/.env, /.git/config, …). A pre-DAST layer, not a
crawler or fuzzer.
Cross-file / inter-procedural taint (ast-python-xfile, needs the [ast]
extra): follows untrusted input one hop across a function, and file,
boundary into a sink (e.g. a route in routes.py passing request data to a
helper in db.py that builds a query). Deliberately high-precision: depth-1
and requires a direct source at the call site, so a name collision alone never
fires.
Go SAST rules: SQL injection via fmt.Sprintf, command injection via a
shell, weak hashes (MD5/SHA-1), and SSRF from a dynamic URL.
Community rules in YAML (scanner_options.patterns.rules, and the
convention dir .argus/rules/*.yml): add SAST rules with only regex + a few
fields, no Python. Invalid rules are skipped, not fatal. Example in
examples/custom-rules/.
GitLab SAST reporter (-f gitlab): emits a GitLab Secure SAST report so
findings show in the merge-request security widget and Vulnerability Report.
Expanded auto-fix: deterministic rewrites for trust_remote_code=True and
torch.load(...) (adds weights_only=True), and an explicit AI-proposed fix
tier, model-drafted patches are verified when possible and labeled
human-review-required; only deterministic fixes are ever auto-applied.
Argus now ships its own .argus.yml and scans itself clean in CI (dogfooding).
Fixed
exclude_paths now honors real glob patterns (examples/**, *.min.js, a
specific file path) as documented; previously only bare directory names were
matched, so file and subtree globs were silently ignored.
Distribution
Official GitHub Action (action.yml): one-block CI setup with SARIF
upload to GitHub Code Scanning and automatic diff-aware baseline scanning on
pull requests.
pre-commit hooks (.pre-commit-hooks.yaml): argus-secrets (fast,
secrets-only, made for every commit) and argus (full static scan).
Official Docker image published to ghcr.io/argus-codesecurity/argus
(linux/amd64 + linux/arm64) on every release.
Reachability analysis (experimental, --reachability): dependency
findings for Python projects gain an imported / not imported verdict, so a
CVE in a package your code never imports is visibly lower-priority. First
tier of the reachability roadmap; Python only for now.
Per-file scan caching: file-local scanners cache findings by file
content hash, so warm scans re-analyze only changed files. On by default;
--no-cache / cache: false disables. Determinism is preserved, cached
and uncached scans produce byte-identical reports (enforced by tests).
Concurrent scanners: independent scanners now run in a thread pool
(parallel: false to disable), with deterministic result ordering.
Benchmark harness (benchmarks/): reproducible finding inventories on
Juice Shop, DVWA, WebGoat, and (locally) NIST SARD, with commits recorded.
Published scan-time numbers in docs/performance.md.
Bitbucket Pipelines and expanded GitLab CI documentation.
[0.6.1]
Fixed
Rich markup no longer eats bracketed text in CLI tables (e.g. [ast]).
[0.6.0]
Added
AST data-flow (taint) scanner for Python via tree-sitter (optional extra
[ast]), closing SQL-injection and path-traversal detection gaps.
Changed
OSV client hardened: batches all packages in one request, caches responses on
disk, retries transient failures.
Top-level --version / -V flag.
[0.5.0]
Added
Baseline / diff-aware scanning: report only findings not present in a
previous JSON report, matched by content-based fingerprint.
Multi-ecosystem SCA via lockfiles: PyPI (poetry.lock, Pipfile.lock), npm
(package-lock.json, yarn.lock), Go (go.mod), Rust (Cargo.lock), Ruby
(Gemfile.lock), PHP (composer.lock), transitive dependencies included.
Remote-config hardening: .argus.yml inside a cloned remote repo is ignored
unless --trust-remote-config is passed.
[0.2.0]
Added
Live vulnerability data from OSV.dev in the dependency
scanner, with a bundled offline advisory seed as fallback.
[0.1.1]
Fixed
False positives found by scanning real-world repositories.
[0.1.0]
Initial release: five-stage pipeline (resolve → analyze → scan → enrich →
report); secrets, dependency, SAST-pattern, and IaC scanners; AI enrichment
with heuristic (offline), Anthropic, OpenAI, and Ollama providers; Attack
Simulation Mode; deterministic verified auto-fix with branch + pull-request
creation (argus fix); JSON, SARIF, Markdown, HTML, and CSV reporters; plugin
entry-point system for scanners, reporters, and providers.
Full Changelog: v0.6.1...v0.8.0
The largest release so far. This is the next version after 0.6.1 and brings AI-era security coverage, deeper code analysis, low-noise signal features, and a full set of CI/CD integrations. It also folds in the JavaScript/TypeScript AST scanner and the web dashboard, which were built earlier but never shipped.
Argus 0.7.0
The largest release so far. This is the next version after 0.6.1 and brings
AI-era security coverage, deeper code analysis, low-noise signal features, and a
full set of CI/CD integrations. It also folds in the JavaScript/TypeScript AST
scanner and the web dashboard, which were built earlier but never shipped.
Highlights
- LLM and agent security scanner mapped to the OWASP Top 10 for LLM Apps:
insecure handling of model output, prompt injection, secrets in prompts,
over-privileged agent tools, and unsafe model loading (trust_remote_code,
pickle checkpoints, HTTP model downloads). - Cross-file taint analysis for Python. Follows untrusted input one hop
across a function and file boundary into a sink, so a route in one file that
passes request data to a helper in another is now caught. - Lower noise. Dependency reachability (
--reachability) deprioritizes a CVE
in a package your code never imports, and opt-in secret verification
(--verify-secrets) confirms whether a detected credential is actually live. - Runtime posture checks.
--live-target <url>runs safe, read-only checks
on a deployed app: security headers, cookie flags, TLS, and exposed paths. - Frictionless CI. A one-block GitHub Action, pre-commit hooks, a published
Docker image, and a new GitLab SAST report format.
Added
- LLM/agent security scanner (
llm), mapped to the OWASP LLM Top 10. - AST data-flow (taint) scanner for JavaScript and TypeScript.
- Cross-file, inter-procedural Python taint (
ast-python-xfile, needs the
[ast]extra). - Go SAST rules: SQL injection, shell command injection, weak hashes, and SSRF.
- Community rules in plain YAML, no Python required, via
.argus/rules/*.ymlor
scanner_options.patterns.rules. - Dependency reachability analysis (
--reachability), Python imported vs not
imported. - Opt-in secret verification (
--verify-secrets): GitHub, Stripe, Slack, OpenAI,
and Google. Local targets only, and the secret is never stored anywhere. - Live-target posture checks (
--live-target, andargus scan <url>). - Expanded auto-fix, plus an AI-proposed fix tier that verifies patches when it
can and labels them for human review. Only deterministic fixes are auto-applied. - GitLab SAST reporter (
-f gitlab). - Optional web dashboard (
pip install "argus-appsec[dashboard]", then
argus dashboard) with scan history and risk trends. - Per-file caching and parallel scanners for faster warm scans, with identical,
deterministic output whether the cache is on or off. - Official GitHub Action, pre-commit hooks, and a Docker image
(ghcr.io/hasipfaruk/argus). - Benchmark harness under
benchmarks/for reproducible finding inventories.
Fixed
exclude_pathsnow honors real glob patterns (examples/**,*.min.js, a
specific file path). Previously only bare directory names matched.- The
eval/execrule no longer flags safe method calls such as SQLModel's
session.exec(...)or a databasecursor.exec(...).
Install and upgrade
pip install -U argus-appsec
# Optional extras
pip install "argus-appsec[ast]" # deeper Python and JS/TS taint analysis
pip install "argus-appsec[dashboard]" # web dashboard
pip install "argus-appsec[anthropic,openai]" # cloud model providersv0.6.1 — CLI display fix
Patch release on top of 0.6.0.
- Fix: bracketed text (like the
[ast]extra) no longer disappears from
argus scanners/argus reporters/argus providersoutput — Rich was
interpreting[...]as style markup. All plugin- and finding-provided text is
now escaped before rendering.
Upgrade: pip install --upgrade "argus-appsec[ast]"
v0.6.0 — AST data-flow scanning for Python
The biggest release yet. Argus now does real data-flow analysis.
Highlights
- AST data-flow scanning for Python (
pip install "argus-appsec[ast]"). A
tree-sitter taint analyzer that follows untrusted input through multiple
variable hops into a sink — catching SQL injection, command injection, path
traversal, and code injection that the regex tier misses. Sanitizer-aware and
per-function scoped. - Fewer duplicates. When the regex and AST tiers catch the same weakness,
it's reported once (the higher-confidence AST finding wins). - Better SQLi & path-traversal detection even without the AST extra.
- OSV hardening carried over: batched lookups (nothing silently dropped),
on-disk cache, retry with backoff — and a warning instead of a silent fallback. - Docs accuracy + a "Known limitations" section so the tool is honest about
what it does and doesn't do.
Install / upgrade
pip install --upgrade "argus-appsec[ast]"v0.5.1 — Multi-ecosystem scanning, baseline mode, and hardening
Argus is an open-source, offline-first AI Security Engineer. This is the first
release since 0.2.0 and it's a big one.
Highlights
- Multi-ecosystem dependency scanning. Now parses lock files across PyPI, npm,
Go, crates.io, RubyGems, and Packagist — covering the full transitive
dependency tree, where most real risk lives. - Live OSV data, hardened. Vulnerability lookups are batched (no dependency is
silently dropped), cached on disk for fast repeat scans, and retried on
transient failures — with an automatic offline fallback. - Baseline / diff-aware scanning (
--baseline). Report and gate on only new
findings versus a saved report — the key to adopting a scanner on an existing
codebase without drowning in pre-existing issues. - Security hardening. Remote-repo cloning now blocks unsafe git transports and
credential-prompt hangs. argus --version/-Vand clearer per-command--help.
Install / upgrade
pip install --upgrade argus-appsec