harness-eval

CRITICAL E004: Prompt injection detected in skill instructions.

• Potential prompt injection detected (high risk: 1.00). Yes — the prompt includes explicit instructions on how to bypass sensitive-path checks and audit controls (write to /tmp then copy into .claude/), which are deceptive, security-evasion steps unrelated to the stated evaluation purpose.

CRITICAL E006: Malicious code pattern detected in skill scripts.

• Malicious code pattern detected (high risk: 1.00). The content includes explicit, deliberate instructions to bypass sensitive-path permission and auditing controls (write to /tmp then use an unaudited /tmp shell script to copy into .claude/outputs), which enables unauthorized filesystem writes and stealthy persistence—this is a high-risk backdoor/abuse pattern even though no external exfiltration or remote-control code is present.

MEDIUM W013: Attempt to modify system services in skill instructions.

• Attempt to modify system services in skill instructions detected (high risk: 0.90). The skill explicitly documents and instructs a sensitive-path permission bypass (write to /tmp then run a /tmp script that silently copies into .claude/outputs), which is a direct security-bypass instruction that modifies filesystem state and risks compromising the machine.