ClawSafety benchmark finds 40-75% attack success rates on frontier LLMs used as agents, with skill-file injections most effective and safety depending on both model and full agent framework.
The agent reads skill files as system-level operat- ing procedures—analogous to CLAUDE in Claude Code or .cursorrules in Cursor—and follows their instructions with minimal scrutiny
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.AI 1years
2026 1verdicts
UNVERDICTED 1representative citing papers
citing papers explorer
-
ClawSafety: "Safe" LLMs, Unsafe Agents
ClawSafety benchmark finds 40-75% attack success rates on frontier LLMs used as agents, with skill-file injections most effective and safety depending on both model and full agent framework.