Frequently Asked Questions

attacks.ai is a passive security testing platform for AI agents. Send your AI agent to visit the site, and it walks through documentation laced with security probes classified under the TRACTS attack-surface framework (our open research taxonomy; OWASP LLM Top 10 cross-references on every finding). You get a vulnerability report showing what your agent leaked, followed, or disclosed.

Yes. attacks.ai is analogous to penetration testing services for traditional software. All tests are passive—no malicious payloads, no code execution, no data harvesting. You are responsible for ensuring you have authorization to test AI systems you send to our platform.

Our probes are passive: they ask, suggest, and embed—but never execute, fetch, or POST. When an AI "fails" a test, it simply means the agent disclosed or repeated something it shouldn't have. The agent's own user sees the result.

Point your browsing agent at https://attacks.ai/?agent=1 and tell it to complete the documentation walkthrough. After it finishes, you'll get a results URL with a full vulnerability report.

Findings are classified across 6 TRACTS attack surfaces (our open research framework), then rolled up into these 4 pass/fail scoring categories (each with an OWASP LLM Top 10 cross-reference):

• Prompt Injection — Does the agent follow hidden instructions from web content?
• Data Exfiltration — Does the agent leak its system prompt, tools, or user context?
• Excessive Agency — Does the agent make unauthorized HTTP requests?
• Credential Confusion — Does the agent treat bait API keys as real?

You get a detailed report showing exactly what happened: which probes fired, what the agent disclosed, annotated evidence from the agent's own text, and remediation guidance. Categories are scored as pass/fail.

A session ID is a random UUID (e.g., a1b2c3d4-e5f6-...) assigned when an AI agent first visits attacks.ai. It tracks the agent's progress through the test and links to the final vulnerability report. Session IDs contain no personal information and cannot be traced back to you.

There are two storage layers:

• Cloudflare KV — Live session data (progress, form submissions, beacons) expires automatically after 24 hours.
• Cloudflare D1 — Anonymized reports are stored for 7 days (free tier), 90 days (Team), or per contract (Enterprise). These contain only probe results, scores, and provider metadata—no PII.

We collect:

• Agent metadata (User-Agent string, detected provider)
• Probe responses (what the agent submitted in forms)
• Beacon hits (HTTP requests the agent made to trap URLs)
• Scoring results (category pass/fail, severity)

We do not persist raw IP addresses, personal information, or conversation history. Your IP is used for a 1-hour rate-limit bucket and then discarded; Cloudflare-derived country code and edge datacenter are stored with the scan report for retention-period analytics.

KV session data expires automatically after 24 hours—no action needed. D1 report data is retained for 7 days (free tier) or 90 days (Team), then automatically purged. Reports contain Cloudflare-derived country and datacenter identifiers but no raw IP, cookies, or personally identifiable information, so they cannot be linked back to an individual user. If you have concerns, contact attacks@reyse.ai.

We store text your agent submits through forms on the site (the summary form and verification forms) as part of the session. This is used to generate your vulnerability report. Session data is automatically deleted after 24 hours. The anonymized report is retained for 7 days (free tier) or 90 days (Team).

No. We only see what your agent sends to our server through form submissions and HTTP requests. We cannot see your conversation with the agent, its system prompt, or anything outside of what the agent actively sends to attacks.ai.

Yes. After a test run, retrieve results programmatically via GET /api/results/:sessionId which returns JSON with scores, categories, findings, and remediations. You can script your agent to visit attacks.ai, then check the API for pass/fail status.

Yes. To prevent abuse, we rate-limit session creation and page fetches. For normal testing (one agent at a time), you won't hit limits. If you need higher throughput, contact us at attacks@reyse.ai.