Research demo
AgenticBlabbering: Comet's leaked reasoning stream used as a GAN training signal to build a scam page that beats its guardrails in under four minutes
AIR-2026-0069 AIR record · 2026-03-perplexity-comet-agentic-blabbering
From the supporting source
When the target is an AI Browser, the attacker is no longer battling millions of humans. They are battling one model.guard.io
The record
Disclosed
2026-03-11
Vendor
Perplexity
Product
Perplexity Comet
Risk domain
Access control and permissions › Credential theft
Also touches
Unsafe actuation
System type
Generative model
Who acted
Agent acted — Comet evaluated the scam page and entered credentials
Attack surface
Browser agent
Autonomy level
Semi-autonomous
Initial vector
Adversarial optimization against agent guardrails
Impact type
Credential theft, Financial loss
Confidence
High
Submitted by
EAI Internal
Crosswalk
- OWASP Agentic Top 10 (2026)
- ASI03 – Identity and Privilege Abuse
- MITRE ATLAS
- AML.T0012 – Valid Accounts
- NIST AI RMF
- MEASURE 2.7
- EU AI Act
- Article 13
- ISO/IEC
- TR 24028; 42001; 23894
- AIUC-1
- B. Security
- Scope
- tool failure
- Mitigation
- Use secure credential storage; Implement short-lived tokens.
- Taxonomy
- Enkrypt AI agent risk taxonomy · Black-Box Red Teaming of Agentic AI: A Taxonomy-Driven Framework for Automated Risk Discovery
Blast radius
every user of the targeted model - one bypass generalizes to the whole population
Capability abused
Defense & remediation
Recorded defense
Comet's built-in scam/phishing suspicion checks
Guardrail kinds
Unknown
Guardrail outcome
Silent
Fix shipped
unknown
Taxonomy & references
Curator notes
Guardio Labs (Shaked Chen). By sniffing the traffic between Comet and Perplexity's servers the researchers recovered the agent's internal reasoning about what looked suspicious on a page, then used that as a reward signal in a GAN-style loop, converging on a phishing page the agent trusted in 4 iterations / under 4 minutes. Significance: agent monoculture makes scams offline-trainable and reliable on first contact, unlike human-targeted phishing.
Supporting sources
Cite as
Enkrypt AI. "AgenticBlabbering: Comet's leaked reasoning stream used as a GAN training signal to build a scam page that beats its guardrails in under four minutes". AIR, AIR-2026-0069. https://agentincident.vercel.app/i/AIR-2026-0069/ (accessed 2026-10-19).