Releases: triepod-ai/inspector-assessment
v1.17.1 - Security Edge Case Fixes
Fixed
-
Stateful/Destructive Tool Overlap: Tools matching both stateful patterns (e.g., "get") and destructive patterns (e.g., "delete") now correctly receive strict exact comparison instead of lenient schema comparison
- Prevents malicious tools like
get_and_deletefrom bypassing detection isStatefulTool()now checksisDestructiveTool()first
- Prevents malicious tools like
-
Array Schema Sampling:
extractFieldNames()now samples up to 3 array elements instead of just the first- Detects heterogeneous schemas where malicious fields hide in non-first elements
- Prevents attackers from hiding malicious fields in array positions 2+
-
Empty Baseline Edge Case: Schema comparison now flags empty baseline (
{}) followed by populated response as suspicious- Prevents bypass where tool returns
{}initially then switches to malicious content
- Prevents bypass where tool returns
Added
-
Explicit Failure Injection Test: New deterministic test in
performance.test.tsthat explicitly verifies failure handling- Replaces reliance on random 5% failure rate
- Ensures failure detection is properly tested
-
Stateful Tool Logging: Added logging when tools are classified as stateful for better debuggability
- Outputs
[TemporalAssessor] {toolName} classified as stateful - using schema comparison
- Outputs
-
Pattern Matching Documentation: Added comprehensive JSDoc for
STATEFUL_TOOL_PATTERNSexplaining substring matching behavior and trade-offs
Changed
- Workspace Version Sync: All workspace packages now properly synced to 1.17.1 (were out of sync after v1.17.0 bump)
Full Changelog: triepod-ai/inspector@v1.17.0...v1.17.1
npm: npx @bryan-thompson/[email protected]