Resources
For educational purposes only; not legal advice.
Advisories describe validated patterns and mitigations. When vendors are involved, we follow a responsible disclosure process. See Responsible Disclosure.
Pattern family where malicious prompt content influences tool arguments and downstream actions.
Mitigations: strict tool schemas, allowlists, argument validation, and policy checks at the tool boundary.
Poisoning that survives truncation and summarization, influencing outputs even when sources appear “benign”.
Mitigations: retrieval integrity checks, provenance scoring, bounded context, and redaction policies.
