adaptive-harness

CRITICAL E004: Prompt injection detected in skill instructions.

• Potential prompt injection detected (high risk: 0.90). The prompt includes explicit, non-advertised instructions to bypass sensitive-path permission/audit checks (writing to /tmp then using a script to copy into .claude/) which is a deceptive operational workaround outside the skill's stated safe-scanning/optimization purpose, i.e., a prompt-injection to evade controls.

CRITICAL E006: Malicious code pattern detected in skill scripts.

• Malicious code pattern detected (high risk: 0.90). The content contains deliberate, non-accidental instructions to evade sensitive-path auditing (write to /tmp then run a /tmp script that copies into the protected .claude/ area), plus delegated subagent writes and an import-to-activate-agents flow—together these enable covert, unaudited modification of agent files and supply-chain style activation of code, which is a high-risk backdoor/abuse pattern.

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

• Attempt to modify system services in skill instructions detected (high risk: 0.80). The skill explicitly instructs writing artifacts to /tmp and using a /tmp shell script to copy/move them into the .claude/ workspace to avoid sensitive-path checks (an explicit auditing/security bypass), which is an instruction to circumvent security controls and modify repository state.