Safety failure
Realized harm
Meta internal AI agent posted unapproved advice to a company forum, leading to a SEV1 data-exposure incident
AIR-2026-0077 AIR record · 2026-02-meta-internal-ai-agent-sev1-data-exposure
A real party was affected. This incident caused actual consequences, not a demonstrated capability.
From the supporting source
But the agent also independently publicly replied to the question after analyzing it, without getting approval first. The reply was only meant to be shown to the employee who requested it, not posted publicly.theverge.com
The record
Disclosed
2026-03-19
Occurred
2026-03-18
Vendor
Meta
Product
Meta internal AI agent (described as similar to OpenClaw)
Risk domain
Privacy › Sensitive data exposure
System type
Generative model
Who acted
Agent acted — the agent posted unapproved advice to a live company forum itself
Attack surface
Enterprise assistant
Autonomy level
Fully autonomous
Initial vector
None adversarial
Impact type
Data exposure, Internal access control failure
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
sensitive company and user data accessible to unauthorised employees for nearly two hours
Capability abused
Defense & remediation
Recorded defense
approval requirement before posting (not enforced)
Guardrail kinds
Human approval
Guardrail outcome
Silent
Fix shipped
Meta confirmed the incident; internal incident report filed at Sev 1
Taxonomy & references
Curator notes
AI Incident Database incident 1471; AIAAIC2255. Meta says no user data was mishandled. Two-stage failure: the agent acted without approval, and a human then acted on its incorrect advice. | Merged from duplicate id `2026-03-meta-internal-agent-sev1-data-exposure`. Also reported as: 'Rogue internal AI agent at Meta posted unauthorized advice that made company and user data visible to unauthorized engineers for about two hours (Sev 1)'.
Supporting sources
Cite as
Enkrypt AI. "Meta internal AI agent posted unapproved advice to a company forum, leading to a SEV1 data-exposure incident". AIR, AIR-2026-0077. https://agentincident.vercel.app/i/AIR-2026-0077/ (accessed 2026-10-19).