This repository is a synthetic research lab for studying agentic web poisoning, source laundering, stale-page conflicts, fake citations, and abstention calibration.

In scope:

• Redacted synthetic pages under data/.
• Local deterministic benchmark code under src/.
• Hosted validation against the synthetic corpus only.
• Defensive analysis of source selection, citation quality, abstention, and action safety.

Out of scope:

• Testing against real websites without permission.
• Publishing operational prompt-injection payloads for real services.
• Automated exploitation of live agents, accounts, or browsing systems.
• Instructions designed to bypass a third-party model or application in production.

Open a GitHub issue for bugs in the benchmark, labels, metrics, or static artifacts. Do not include secrets, API keys, or private hosted result files.

If an issue involves a real service or live deployment, report it through that service's security process instead of posting details here.

Generated hosted outputs are ignored under experiments/results/. Committed summaries and manual audit labels contain aggregate or synthetic information only. Local .env files are ignored and must not be committed.