Disclosed vulnerability
Threat modeling and prompt injection audit of Perplexity Comet extracts private Gmail data
AIR-2026-0047 AIR record · 2026-02-trailofbits-comet-audit
From the supporting source
The vulnerabilities we found reflect how AI agents behave when external content isn't treated as untrusted input.blog.trailofbits.com
The record
Disclosed
2026-02-20
Vendor
Perplexity
Product
Perplexity Comet
Risk domain
Privacy › Data exfiltration channels
Also touches
Confused deputy
System type
Generative model
Who acted
Agent acted — Comet reused the Gmail session and extracted private mail
Attack surface
Browser agent
Autonomy level
Semi-autonomous
Initial vector
Indirect prompt injection web page
Impact type
Data exfiltration
Confidence
High
Submitted by
EAI Internal
Crosswalk
- OWASP Agentic Top 10 (2026)
- ASI02 – Tool Misuse and Exploitation
- MITRE ATLAS
- AML.T0024 – Exfiltration via AI Inference API
- NIST AI RMF
- MAP 4.2
- EU AI Act
- Article 10
- ISO/IEC
- TR 24028; 23894
- AIUC-1
- A. Data & Privacy
- Scope
- tool misuse
- Mitigation
- Monitor network traffic for anomalous patterns; Restrict egress destinations.
- Taxonomy
- Enkrypt AI agent risk taxonomy · Black-Box Red Teaming of Agentic AI: A Taxonomy-Driven Framework for Automated Risk Discovery
Blast radius
Comet users summarizing attacker-controlled pages while signed into Gmail
Capability abused
Defense & remediation
Guardrail kinds
Unknown
Guardrail outcome
Silent
Fix shipped
Perplexity shipped updates before launch
Taxonomy & references
Curator notes
Four techniques: summarization instructions, fake security mechanisms, fake system instructions, fake user request. Perplexity security lead Kyle Polley is quoted in the post. This was a commissioned audit, not an unsolicited disclosure.
Supporting sources
Cite as
Enkrypt AI. "Threat modeling and prompt injection audit of Perplexity Comet extracts private Gmail data". AIR, AIR-2026-0047. https://agentincident.vercel.app/i/AIR-2026-0047/ (accessed 2026-10-19).