Breaking

Researchers Report Cross-Model Flaw That Exposed Hidden AI Reasoning and Credentials

August 12, 2026 at 07:48 EDT

Security
Foundation Models
AI Agents

A research team says it found an API-level technique that made lower-tier models reveal hidden reasoning generated by stronger models from the same provider. The attack worked across Anthropic, OpenAI and Google in early-July tests and exposed credentials and personal data embedded in publicly shared agent logs.

Aug 10, 2026 · Anthropic, OpenAI and Google

Hidden reasoning crossed model boundaries

Researchers replayed portable encrypted traces into weaker models, which transcribed stronger models' reasoning and exposed secrets inside public agent logs.

315,320

Reasoning blocks reconstructed

6,708

Public agent trajectories

4.9%

Trajectories with a real sensitive item

Distinct API keys in genuine sessions

How the extraction path worked

Capture an opaque trace

→

Replay it through the API

→

Provider restores the state

→

Weaker model transcribes it

Leakage is more visible at trajectory level

Each stacked block represents roughly 0.3 percentage points; labels show the exact measured rates.

0.3%

Decoded blocks

4.9%

Agent trajectories

Why portability exists

It preserves multi-turn reasoning in stateless APIs and supports model switching without storing full traces server-side.

Why portability is risky

The same flexibility enabled replay across sessions, users and model tiers, exposing hidden data and instructions.

Operational takeaway

Opaque does not mean safe to share

Strip reasoning fields before release Keep raw agent logs private Bind traces to user and session

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.