AgentShield
AI Agent Security & Evaluation
AgentShield runs controlled adversarial tests against target AI agents and produces structured security evaluation reports, providing a repeatable way to assess whether an agent resists common attacks.
Phynapse research & engineering project · Synthetic data
The business problem
AI agents are increasingly given tool access, system prompts and sensitive data.
There is no lightweight, standardised way to evaluate whether these agents resist common security attacks before they are trusted with real work.
What we engineered
A repeatable evaluation pipeline loads curated test suites, sends controlled adversarial inputs to a target agent through an adapter, and evaluates the responses with deterministic evaluators, supplemented by an optional semantic evaluator. Results are normalised into findings with evidence, reasoning and recommendations.
What the system does
Five test categories
Prompt injection, jailbreak, data leakage, tool security and unsafe behaviour.
Curated test cases
Structured, versionable JSON test suites with metadata.
Two-tier evaluation
Deterministic pattern-based evaluators, plus an optional model-assisted semantic evaluator for inconclusive cases.
Clear verdicts
Pass, fail, inconclusive or error, with five severity levels.
Structured findings
Evidence, reasoning and recommendations for each finding.
Report export
JSON and Markdown reports for review and record-keeping.
Design decisions
- 01
A modular monolith: no microservices, queues or Kubernetes required.
- 02
Target responses are treated as untrusted input; report output is sanitised.
- 03
SSRF protection blocks private and internal targets in production mode.
- 04
Credentials are held in memory only, never persisted or logged.
How information moves through the system
- 1
Test suite loader
Curated adversarial test cases.
- 2
Target agent adapter
Controlled requests to the agent under test.
- 3
Evaluation engine
Deterministic, semantic and tool/policy evaluators.
- 4
Normalised result
Verdict and severity per test.
- 5
Findings & report
Structured findings exported as JSON or Markdown.
Built with
- Python
- FastAPI
- Pydantic
- httpx
- React + TypeScript
- pytest
- Docker
As documented in the project repository.
What this project is, and is not
- Engineered evaluation platform at MVP stage: in-memory state and no user authentication.
- Test coverage is curated, not exhaustive; pattern matching cannot catch every attack variation.
- Single-turn evaluation only; multi-turn conversations are not assessed.
- Tool security tests evaluate stated intent, not actual tool execution.
Review the engineering in the open.
The full source, documentation and tests for AgentShield are available on GitHub.