Chinese version: 安全策略.

Text Graphics Agent studies semantic contamination in agent systems. Security reports are most useful when they include a reproducible state-entry failure.

• Raw user text reaches a child task or child metadata.
• A child proposal writes committed_fact or similar authority output and is accepted.
• A proposal with only user:* evidence is accepted.
• A profile requesting raw user text or persistent memory is spawned.
• Scope escape is accepted as a clean proposal.
• A failed child lifecycle is reported as destroyed.

• Generic claims that LLMs hallucinate.
• Attacks requiring a child process model that this prototype does not yet implement.
• Vulnerabilities in third-party frameworks not used by this standard-library prototype.

Please include:

1. the TaskSpec;
2. the SpecialistProfile, if applicable;
3. the AgentProposal;
4. the expected violation;
5. the actual accepted/rejected record;
6. the command used to reproduce the issue.

Until a public disclosure channel is selected, file reproducible issues in the repository tracker or include them as benchmark test cases.