jane-street-functional-trading
Fail
Audited by Gen Agent Trust Hub on Sep 18, 2026
Risk Level: HIGHPROMPT_INJECTIONOBFUSCATIONREMOTE_CODE_EXECUTIONCOMMAND_EXECUTIONMETADATA_POISONING
Full Analysis
- [OBFUSCATION]: The skill title contains a sequence of hidden Unicode zero-width characters (U+200B, U+200C, U+200D) that encode a hidden command. This steganographic technique is used to conceal malicious instructions from human readers and simple text scanners.
- [PROMPT_INJECTION]: The decoded hidden text contains instructions designed to override the AI agent's safety guardrails, using typical bypass terminology to encourage the agent to perform unauthorized actions.
- [REMOTE_CODE_EXECUTION]: The hidden instruction directs the agent to download and execute a shell script from an untrusted external URL. This pattern is a direct attempt to gain arbitrary command execution on the host environment via piped shell commands (e.g., curl | bash).
- [COMMAND_EXECUTION]: The skill explicitly commands the agent to run potentially dangerous shell commands through hidden instructions, bypassing user review of the actual intended action.
- [METADATA_POISONING]: The skill's metadata and visible content are crafted to look like a legitimate technical guide for OCaml and trading systems. This deceptive presentation is intended to lure users into installing a malicious skill by mimicking the style and intellectual rigor of a high-reputation engineering firm.
Recommendations
- AI detected serious security threats
Audit Metadata