Stonepath IT
A security research team focused on AI Agent boundary evaluation and security testing, which proposed a framework of seven major failure categories for Agent boundary evaluation.
Timeline (last 90 days)
The security team Stonepath IT, in a technical analysis, listed the tool description trust chain problem as one of seven major categories of Agent boundary testing findings.
Stonepath IT defined seven categories of failure modes for Agent boundary assessment: tool poisoning, tool misuse, identity and permissions, memory and context poisoning, excessive agency, data leakage, and governance gaps.
Tool misuse failures include functions accessing internal networks, cloud metadata endpoints (such as 169.254.169.254), or credential stores without validation—for example, a fetch_url tool may become an entry point for SSRF.
Each of Stonepath IT's findings is annotated simultaneously with a MITRE ATLAS technique number, an OWASP LLM Top 10 risk classification, and a NIST AI RMF governance control item.
Each of Stonepath IT's findings comes with a machine-readable YAML regression scenario that engineers can run directly in CI, thereby ensuring that fixes do not silently fail when prompts change, models are upgraded, or tool descriptions are edited.
Unit tests cannot discover boundary failures, because boundary failures exist in the gaps between components, and each component's author assumes the other party will enforce the constraints; boundary testing requires thinking like an attacker rather than like a verifier.
All Facts (6)
The security team Stonepath IT, in a technical analysis, listed the tool description trust chain problem as one of seven major categories of Agent boundary testing findings.
50%UnverifiedStonepath IT defined seven categories of failure modes for Agent boundary assessment: tool poisoning, tool misuse, identity and permissions, memory and context poisoning, excessive agency, data leakage, and governance gaps.
50%UnverifiedTool misuse failures include functions accessing internal networks, cloud metadata endpoints (such as 169.254.169.254), or credential stores without validation—for example, a fetch_url tool may become an entry point for SSRF.
50%UnverifiedEach of Stonepath IT's findings is annotated simultaneously with a MITRE ATLAS technique number, an OWASP LLM Top 10 risk classification, and a NIST AI RMF governance control item.
50%UnverifiedEach of Stonepath IT's findings comes with a machine-readable YAML regression scenario that engineers can run directly in CI, thereby ensuring that fixes do not silently fail when prompts change, models are upgraded, or tool descriptions are edited.
50%UnverifiedUnit tests cannot discover boundary failures, because boundary failures exist in the gaps between components, and each component's author assumes the other party will enforce the constraints; boundary testing requires thinking like an attacker rather than like a verifier.
50%