fix(middleware): improve pre-tool middleware guarding logic#1824

Merged

rapids-bot[bot] merged 24 commits intoNVIDIA:developfrom

cparadis-nvidia:fix-pre-tool-verifier-defense-middleware

Apr 8, 2026

Commits on Mar 30, 2026

security fix: _analyze_content now splits long content into sequential chunks and verifies each one, with _analyze_chunk handling the
cparadis-nvidia
committed
new test file covering chunking mechanics (including the specific attack scenario that was previously exploitable), all three
cparadis-nvidia
committed
uses a sliding window with _STRIDE = _MAX_CONTENT_LENGTH // 2
cparadis-nvidia
committed
updated unit tests
cparadis-nvidia
committed
At most _MAX_CHUNKS windows are analyzed.
cparadis-nvidia
committed
updated unit tests
cparadis-nvidia
committed
json.dumps handles all fields correctly
cparadis-nvidia
committed
the label text notes the content is HTML-escaped so the verifier treats tags as literal characters
cparadis-nvidia
committed
updated unit tests
cparadis-nvidia
committed
over-cap inputs are randomly sampled and still analyzed (up to _MAX_CHUNKS calls)
cparadis-nvidia
committed

Commits on Apr 2, 2026

Commits on Apr 3, 2026

Merge branch 'develop' of github.com:NVIDIA/NeMo-Agent-Toolkit into fix-pre-tool-verifier-defense-middleware
cparadis-nvidia
committed

Commits on Apr 7, 2026