You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
This commit was created on GitHub.com and signed with GitHub’s verified signature.
Added
Test-verified fixes.nephilex fix --verify-with-tests (with --test-command, or autofix.test_command from 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-selection and llm.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, so scan and check report the same rule
and check keeps 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
--format help on scan and import now lists junit.
Documentation that did not match the code: the number of report formats, --live-target described as checking TLS (it checks HTTPS and the redirect),
and dependency reachability and VEX not_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
a RecursionError. 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.