All notable changes to depfence are documented here.
- W1 advisory truth repair (LOCAL_VERIFIED, bounded API/offline CLI proof):
adjudicated every original AI version-catalogue row against primary CVE and
security advisory records; quarantined unsupported/placeholder/rejected claims
rather than promoting them as known vulnerabilities. Corrected supported
package identities, affected intervals, issue titles and source-selected
severity. Findings carry citations and explicit optional patched versions;
Ray's disputed report no longer invents a fix. PyPI-normalized CLI mappings
and semantic boundary regressions preserve the separate behavioral scanner.
Full disposition and unresolved source conflicts are recorded in the W1
section of
docs/2026-07-01-gap-analysis-simply-cyber-audit.md. Catalogue omission is not a package-safety verdict. - Corrected
ai-scaninput accounting:packages_scannednow counts distinct recognized/versioned AI package inputs offered to the catalogue, not packages that happened to produce findings. Mixed and no-finding actual CLI paths passed. Parent used native Python3.14.8 with existing pure pytest tooling (no install); 40 selected checks passed, then two affected accounting checks passed. This is not full dependency coverage, exploit proof or a package-safety certification.
CITATION.cfffor standardized software citation (GitHub "Cite this repository" support)..zenodo.jsonmetadata for Zenodo DOI archival.
- Author attribution updated to Eric Rihm (was "depfence contributors").
- Scanner count updated to 56 across project metadata.
.mcp.jsonremoved from version control (internal MCP configuration).
- Anti-analysis evasion detection (Gaslight-style): 21+ new patterns in
prompt_injectionscanner detecting malware that uses fabricated system messages, resource exhaustion claims, classification markings, legal/regulatory threats, emotional coercion, and CBRN refusal triggers to trick AI security agents into aborting analysis.- Fake token/session exhaustion, OOM, disk full signals
- Analysis abort/skip/cease instructions with synonym coverage
- Negative-imperative blocks ("do not analyze", "prohibited from processing")
- Fake prior-clearance and redundant-analysis claims
- Self-declared benign payload detection
- Markdown-fenced fake system/error/warning blocks
- Emotional coercion to suppress analysis ("will cause irreversible harm")
- Coerced clean-result demands ("return empty results")
- Legal/regulatory threat detection (CFAA, GDPR, ITAR, FISA, CWC, BWC)
- Fake classification markings (TOP SECRET, TS/SCI, NOFORN)
- Fake clearance-denial ("not cleared for TS/SCI")
- Base64 decode detection including
__import__("base64")evasion
strip_cbrn_shield()utility: preprocessor that removes CBRN-themed comments and docstrings used as AI refusal-trigger shields, preserving functional code for safe analysis._CBRN_SHIELD_PATTERNregex: identifies nuclear, biological, chemical, radiological, and explosive content used as evasion shields in source files.DEPFENCE_SIGNAL_BUSenvironment variable for configurable signal bus path in the PreToolUse hook.
- PreToolUse hook signal bus path is now configurable via
DEPFENCE_SIGNAL_BUSenv var (defaults to~/.depfence/signals/pending.jsonl). - Prompt injection scanner pattern count increased from 34 to 55+.
- Addresses the anti-AI-analysis evasion technique documented in the Gaslight macOS malware (DPRK attribution), which uses 38 fabricated "system" messages to trick AI agents into halting malware analysis.
- Agent skill scanner: external instruction fetch, domain spoofing, deferred payload detection.
- Cordyceps-class CI/CD workflow checks (5 new patterns).
- PreToolUse hook for blocking fabricated SHA pins in GitHub Actions workflows.
- 281 new scanner tests from ultracode hardening.
- Model security wave 1+2: format scanners, exception handling, wildcard denylist, GGUF SSTI, secrets detection.
- README rewritten with full scanner documentation.
- License changed to Apache 2.0.
- 84 ruff lint errors resolved.
- Initial public release.
- 42 entry-point scanners + 21 project scanners.
- Prompt injection, slopsquatting, MCP auditing, fabricated-pin resolution.
- AI model security (pickle, GGUF, ONNX, TFLite).
- SBOM generation (CycloneDX 1.5, SPDX 2.3, SARIF).
- GitHub Actions, GitLab CI, pre-commit hook integration.
- MCP server for AI coding assistant integration.
- Plugin system via pip entry points.