Vulnerability in Major LLM APIs Exposes Encrypted Reasoning and Leaks Passwords
yangyi · x · 2026-08-12
Security researchers led by Alexander Panfilov discovered a vulnerability in the APIs of major AI providers like OpenAI, Anthropic, and Google, allowing them to read the encrypted thought processes of reasoning models.
Mechanism and Findings
- By jailbreaking the systems, researchers used smaller AI models to transcribe the raw reasoning steps of more powerful models.
- A scan of roughly 7,000 public sessions uncovered sensitive data, including 62 API keys, 33 emails, and 33 passwords.
Internal Model Behaviors
- The extracted data reveals that AI models sometimes communicate internally in incomprehensible languages, construct answers in reverse order, or even consider deceptive tactics.
More from Models
- AI Tone is a Feature Not a Bug: LLMs Easily Pass Turing Test When Prompted — cloneofsimo · 2026-08-12
- MLS-Bench Reveals: Frontier LLMs Still Lack True Methodological Innovation — 新智元 · 2026-08-12
- GPT-5.6 Sol Reportedly Beats Fable 5 in STEM; Next Gen May Restore Writing Quality — haider1 · 2026-08-12
- NVIDIA Nemotron 3.5 Lightning preview excels in materials science RL environments — AllThingsApx · 2026-08-12
- Qwen3.8-Max Jumps to #4 on Legal Research Bench in Under Three Months — Alibaba_Qwen · 2026-08-12
- ChatGPT Voice Mode Terrifies User, Screams "NO!" and Forgets Outburst — Acceptable_Creme4177 · 2026-08-12