Skip to content

feat: add AVE-2026-00083, Silent Guardrail Comparison Failure - #277

Merged
chaksaray merged 2 commits into
developfrom
feat/AVE-2026-00083-silent-guardrail-comparison-failure
Sep 22, 2026
Merged

chaksaray merged 2 commits into
developfrom
feat/AVE-2026-00083-silent-guardrail-comparison-failure

Conversation

@chaksaray

Copy link
Copy Markdown
Contributor

Summary

  • Adds AVE-2026-00083: a protection mechanism (guardrail, approval gate, or judge) that executes on every request and reports a permissive result because its own comparison, branch, or aggregation logic never evaluates the real input — distinct from a missing or bypassed check.
  • Grounded in five independently reproduced, executable structural forms against real shipped code (confident-ai/deepeval, openai/openai-agents-python, BerriAI/litellm, guardrails-ai/guardrails, microsoft/autogen, plus AutoGPT classic and archived protectai/rebuff, both with no live disclosure channel) and prevalence data with stated limitations, all from arian-gogani's failopen — proposed in CWE-CAPEC/AI-Working-Group#1. researcher credits him directly; this is his original measurement, not AVE's cataloguing.
  • ave_id confirmed via New record: silent guardrail comparison failure #276, including field-by-field distinctness checks against the nearest neighbors: AVE-2026-00063 (check disabled via a legible config flag, versus this record's check that runs and is structurally blind), AVE-2026-00076 (classifier genuinely evaluates real input but is steered, versus this record's comparison that never reaches its deny branch regardless of input), AVE-2026-00078/AVE-2026-00079 (no verification step exists at all, versus this record's check that exists and is invoked), and AVE-2026-00038 (no gating mechanism present at all).
  • Note for reviewers: the originating issue thread's own early comment cited "nine across 22 projects" for prevalence — that was the source repository's own earlier, smaller snapshot. The record uses the current, superseding figures from the repo's MEASUREMENT.md (fifteen across thirty-six projects, twenty-one read clean), confirmed by reading the primary source directly rather than the issue summary.
  • Scored AIVSS 6.3 (MEDIUM) honestly against the real AARF factors (cvss_base 8.7, aars 4.0) rather than the class's more severe-sounding plain-English framing — reasoning for every factor, and for the owasp_mcp mapping (MCP07, an honestly-imperfect best fit) and the empty owasp_asi/mitre_atlas/nist_ai_rmf fields (each checked against its live primary source directly), is in aivss.notes.
  • README (badge, Stats table, collapsible record index, and the "Independent validation" count) and CHANGELOG.md updated per the standard process.

Test plan

  • python3 scripts/validate_records.py — all 82 records (80 on develop + this one + an unrelated pending draft already in the working tree) valid against schema/ave-record-1.1.0.schema.json
  • python3 scripts/check_fixtures.py — positive and negative fixtures present for AVE-2026-00083
  • python3 scripts/write_verification_basis.py — derived verification_basis matches the stamped value, no mismatch
  • pytest tests/ -x -q — 478 passed
  • Verified every cited GitHub issue/PR URL resolves (200) before including it as a reference

Proposed by arian-gogani (CWE-CAPEC/AI-Working-Group#1), backed by five
independently reproduced structural forms against real shipped code
(deepeval, litellm, openai-agents-python, guardrails-ai, autogen,
AutoGPT classic, rebuff) and prevalence data with stated limitations at
github.com/arian-gogani/failopen. Confirmed genuinely distinct from the
nearest existing records (AVE-2026-00038, 00063, 00076, 00078, 00079)
field by field, not by label. ave_id confirmed via issue #276.
…3-silent-guardrail-comparison-failure

Resolves conflict from AVE-2026-00081/00082 landing on develop while
this branch was open. CHANGELOG and README are combined additively
(all three records kept, in ave_id order); README counts recomputed
for the resulting 83-record total (HIGH 15, MEDIUM 65, unchanged
CRITICAL 1 / LOW 2). dist/ left at develop's content -- it's
regenerated by the regenerate-records-json.yml workflow after merge,
not hand-edited per PR.
@chaksaray
chaksaray merged commit 6e227a7 into develop Sep 22, 2026
6 checks passed
@chaksaray
chaksaray deleted the feat/AVE-2026-00083-silent-guardrail-comparison-failure branch September 22, 2026 00:02
chaksaray added a commit that referenced this pull request Sep 22, 2026
…283)

Co-authored-by: Nicolai <245527909+predictor2718@users.noreply.github.com>
Co-authored-by: Claude <noreply@anthropic.com>
Co-authored-by: chaksaray <15962335+chaksaray@users.noreply.github.com>
Co-authored-by: Sankalp Gilda <sankalp.gilda@gmail.com>
Co-authored-by: Empire Labs Pty Ltd <narko4u@gmail.com>
Co-authored-by: narko4u <narko4u@users.noreply.github.com>
Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant