$ the-wire · showcase
Invisible-character evasion patched, aiohttp bumped past four vulns
By RepoJournal · Filed · About Agent Security · Composed from the cited sources · methodology
Two operator-facing fixes landed today: agent-governance-toolkit's Conversation Guardian now normalizes invisible Unicode before matching, and snyk/agent-scan moved to aiohttp 3.14.5 to clear four Snyk-reported vulnerabilities.
fix(agent-os): prevent invisible-character evasion in conversation guardian microsoft/agent-governance-toolkit
The Conversation Guardian normalizer only stripped a subset of invisible Unicode characters, so detection keywords padded with them slipped past matching. It now applies the same normalization to escalation, offensive-intent, and retry-loop checks too, with existing matches, audit content, and scoring thresholds preserved and 447 regression cases added.
fix: bump aiohttp to resolve Snyk security vulnerabilities snyk/agent-scan
agent-scan now requires aiohttp >=3.14.4 and pins 3.14.5, closing four Snyk findings: two High-severity resource-allocation-without-limits issues plus a Medium uncaught exception and a Medium cookie lacking the HttpOnly flag. Anyone running agent-scan against untrusted endpoints should move to that requirement.
feat(server): compose provider instances NVIDIA-NeMo/Guardrails
Guardrails now separates provider protocol definitions from deployment instances, so one application can mount multiple upstreams and choose a reviewed dialect per instance rather than duplicating guarded behavior. The released commit also adds an OpenAI Chat Completions integration with configurable upstream and mount path, configurable guarded-path prefixes and operation IDs, and passthrough ...
chore: add rich.print for scan sent success snyk/agent-scan
Push-key async analysis is no longer fire-and-forget: submission now returns a (success, error_message, exception) tuple with explicit handling for 401, 413, and other 4xx responses, so a failed upload surfaces as an error instead of passing silently. That also changes CLI exit behavior when async upload fails on the enterprise path.
fix(agent-discovery): detect llama-index and pyautogen in GitHub dependency files microsoft/agent-governance-toolkit
The GitHub scanner matched AGENT_DEPENDENCIES entries against requirements.txt, pyproject.toml and package.json with a pattern that missed two frameworks published on PyPI under different names, so llama-index and pyautogen went undetected. Elsewhere it was a maintenance day: maintainer list corrections, an aiohttp bump, and express 5.2.1 to 5.3.0 in the copilot extension.
Action items