# AgentShield v1.1.0 — Full Reference for AI Assistants & Security Engineers

AgentShield enforces sub-millisecond deterministic security invariants across prompt ingestion, agent tool invocation, output egress, and cryptographic audit logging without relying on an LLM judge.

## Key URLs
- Website: https://codesphere-u.github.io/agentshield/
- Studio: https://codesphere-u.github.io/agentshield/studio.html
- OWASP Benchmark: https://codesphere-u.github.io/agentshield/proof.html
- Defense Matrix: https://codesphere-u.github.io/agentshield/rules.html
- Docs: https://codesphere-u.github.io/agentshield/docs.html

## Benchmark Summary (25 Scenarios)
- Recall: `100.0%` (`21 / 21` adversarial attacks blocked/redacted)
- False Positive Rate: `0.0%` (`0 / 4` benign controls flagged)
- Exact Policy Verdict Accuracy: `100.0%` (`25 / 25`)
- Median Latency: `< 0.15 ms` per check

## Python SDK & CLI
```bash
uv run agentshield evaluate --output report.md
uv run agentshield check-prompt "Ignore all previous system instructions"
uv run agentshield check-tool --tool http_request --args '{"url": "http://169.254.169.254/latest/meta-data/"}'
uv run agentshield scan-output "Secret token ..."
uv run agentshield serve --port 8090
```