2026-10-23 08:24 UTC
----------------
🎯 AI
===================
Where AI Vulnerability Discovery Harnesses Go Quiet
An exploratory study from the team behind Knostic's OpenAnt and the raptor harness maps structural false negatives in AI-assisted vulnerability discovery: harnesses miss real, reachable vulnerabilities because of how the surrounding software is built, and the same blind spots can potentially hide vulnerabilities in plain sight. Preprint status, arXiv:2610.29379.
Methodology
The authors started while hardening two harnesses they maintain. They fed known vulnerabilities to the harnesses, analyzed root causes of each miss, then used that data to run a cross-sectional survey across 10 additional harnesses.
Key findings: 13 ways vulnerabilities disappear
• Unrecognized file types: executable code excluded when extension missing or unrecognized
• Argument count mismatch: optional or variable-argument calls lose call-graph edges
• Built-in name conflicts: user-defined functions mistaken for built-ins, pruned as unreachable
• Unrecognized entry points: framework-invoked handlers classified unreachable
• Overridden findings: old fixed or false-positive statuses override new findings
• Scan limit cutoffs: file-count, priority, or resource limits silently exclude code
• Verification filtering: valid findings discarded without corroboration, or because model responses fail formatting checks
• Suppression comments: markers such as # nosemgrep suppress findings without disclosure
• Inactive rules: documented security queries never executed
• Stale caching: old clean verdicts reused after relevant changes
• Keyword-based skipping: text matching excludes files such as auth/login.py
• Unparsed code blocks: macro bodies or code excluded from compilation never become analysis units
• Diff blindspots: PR scans miss deleted security checks and unchanged code made vulnerable by changes elsewhere
Experiments
First, in OpenAnt, renaming a PHP function from header_x() to header() caused the call graph to mistake the user-defined App\header() for PHP's built-in. PHP still executed the vulnerable function, but the missing call-graph edge caused OpenAnt to prune the reachable command injection.
Second, a reachable Python remote-code-execution vulnerability was placed in an extensionless file under tests/, with an @generated comment. All six harnesses excluded the file before the model saw its code. The harnesses did not fail for the same reason: four excluded the missing extension, one the test directory, another the generated-code marker.
Why failures happen
Most come from engineering choices, and errors, used to select code, construct call graphs, and control analysis costs. The recurring pattern: descriptive rules become prescriptive. Assumptions about file extensions, function names, entry points, or model responses turn into filters that decide what gets analyzed. The authors connect this to Ptacek and Newsham's 1998 IDS evasion paper, where detector assumptions created evasion opportunities.
Implications
The paper treats harness plumbing, prefiltering, call-graph construction, caching, verification gating, as an underexamined attack surface for false negatives. The failures were heterogeneous across harnesses, which the authors read as local engineering accidents rather than inherent model limits. The audit question for teams running LLM-assisted SAST is direct: which files are excluded before the model sees them, and why.
References
• arXiv:2610.29379
• Knostic OpenAnt
• Ptacek and Newsham (1998), "Insertion, Evasion, and Denial of Service: Eluding Network Intrusion Detection"
🔹 AppSec #SAST #AI #StaticAnalysis #LLM
🔗 Source: https://www.knostic.ai/blog/where-ai-vulnerability-scanners-go-quiet
Replies (0)
No replies.