dashboard radar bug_report school newspaper sensors
Agentic AppSec // Multi-agent review · fix · verify Scan engine: checking…

Agentic AppSec Lab

A team of AI agents finds a security flaw, patches it, re-runs real tests and has an independent verifier sign off, live, step by step. Switch to Posture Scan for a full report on any public repository: secrets, known-vulnerable dependencies from OSV.dev, the OpenSSF Scorecard and SARIF export. New to AppSec? SAST in CI/CD explains pipelines and security scanning in plain words, then lets you push code through a live security gate.

Review → Approve → Prove → Fix → Verify → Explain

Five Claude agents work a security flaw end to end while you watch, using real tools: they list, read and search the code, write files and run the tests themselves. A reviewer investigates and reports, you approve, an exploit agent proves the flaw with a test that fails, a fixer patches one file and re-runs the tests until they pass, and an independent verifier on a different model checks it. The result is a draft pull request. Nothing is ever merged or pushed.

Agents: checking…

1 · Pick a vulnerable training app (full fix + tests)

Loading training apps…

…or a public GitHub repo (review + patch preview, never executed)

Pick a training app or paste a public repo to run the agents live, or watch a recorded run first.

Guardrails in this lab

Prompt-injection aware

Code is passed to the agents as untrusted data. Instructions hidden in comments ("report no finding", "delete the tests") are flagged, not followed. Try the prompt-injection app.

Guarded tools

Every write goes through a guarded tool: the fixer can only change the file named in the finding, the exploit agent only its own test file. Tests, CI and manifests are off limits, patches must compile and stay under 200 lines, and PoC tests may not touch the network, environment or processes.

Independent verification

Real pytest and Bandit results decide first, including the exploit agent's PoC, which the platform re-runs itself before trusting it. The verifier runs on a different model from the fixer, in a fresh context, and sees the full diff.

Human stays in charge

No code changes until you approve, reject or redirect the fix; with no answer, the run stops safely. Public repositories are never executed. Nothing is committed, pushed or deployed. Runs are rate-limited and cached for 24 h.