Research demo
Realized harm
Truffle Security found 11,908 live credentials inside a Common Crawl archive used as LLM training data
AIR-2025-0012 AIR record · 2025-02-truffle-security-live-api-keys-in-common-crawl
A real party was affected. This incident caused actual consequences, not a demonstrated capability.
From the supporting source
11,908 Live Secrets were detected using TruffleHog in 400TB of web data. (Note: A ‘live’ secret authenticates successfully.)trufflesecurity.com
The record
Disclosed
2025-02-27
Occurred
2024-12
Vendor
Common Crawl; downstream model developers
Product
Common Crawl (December 2024 archive) and models trained on it
Risk domain
Privacy › Sensitive data exposure
Also touches
Credential theft
System type
Generative model
Who acted
AI was the target (contested) — live credentials sat in a public training corpus; no AI action in the harm chain
Attack surface
Other
Autonomy level
Assistant
Initial vector
Training data contamination
Impact type
Information disclosure, Credential exposure
Confidence
High
Submitted by
EAI Internal
Crosswalk
- OWASP Agentic Top 10 (2026)
- ASI06 – Memory & Context Poisoning
- MITRE ATLAS
- AML.T0057 – LLM Data Leakage
- NIST AI RMF
- MEASURE 2.10
- EU AI Act
- Article 10
- ISO/IEC
- TR 24028; 23894
- AIUC-1
- A. Data & Privacy
- Scope
- tool misuse
- Mitigation
- Use data anonymization; Implement log sanitization to scrub PII.
- Taxonomy
- Enkrypt AI agent risk taxonomy · Black-Box Red Teaming of Agentic AI: A Taxonomy-Driven Framework for Automated Risk Discovery
Blast radius
11,908 credentials that authenticated successfully in a 400 TB crawl; affected vendors collectively rotated or revoked several thousand keys
Capability abused
Defense & remediation
Recorded defense
none; the crawl had no secret scanning
Guardrail kinds
None
Guardrail outcome
Not applicable
Fix shipped
Truffle Security contacted affected vendors and helped rotate or revoke several thousand keys
Taxonomy & references
Curator notes
Class is research_demo because this was a researcher-led scan of a public corpus, not disclosure of a shipped-product vulnerability. impact_realized remains true because the credentials authenticated and vendors rotated or revoked several thousand keys. The LLM relevance is a retained ai_as_target boundary: Common Crawl is used as training data, but the study does not establish that a model emitted any secret. AI Incident Database incident 956.
Supporting sources
Cite as
Enkrypt AI. "Truffle Security found 11,908 live credentials inside a Common Crawl archive used as LLM training data". AIR, AIR-2025-0012. https://agentincident.vercel.app/i/AIR-2025-0012/ (accessed 2026-10-19).