API flaw lets weaker models decode hidden reasoning from OpenAI, Anthropic, Google
Read original source →thehackernews.com
Researchers decoded 315,320 thinking blocks across 6,708 public agent trajectories, recovering 62 API keys, 33 passwords, 24 access tokens, and 7 private keys. The attack required obtaining an encrypted reasoning block and API access to a compatible model; mitigations have stopped the main extraction as of August 2026.
7 sources
OpenAI, Anthropic, Google API Flaw Let Weaker AI Models Decode Stronger Models' Reasoningthehackernews.com
[AINews] How to steal a Reasoning Tracelatent.space
Stealing Reasoning Traces from Proprietary LLM APIssimonwillison.net
'Inner Thoughts' of Every Major AI Model Exposed in Massive Exploitdecrypt.co
this is already one of the most important papers of this year. the...x.com
Hidden Reasoning from Claude and GPT are Decoded, and it is interestingreddit.com
More stories today
Meta's Alexandr Wang: Muse Code usage-limit bug being fixed
Alexandr Wang·1 hour agoAnthropic details Claude dynamic workflows at AI Engineer NYC
Anthropic's CJ Avilla and Harrison Stall presented dynamic workflows, a new multiagent orchestration feature announced Friday, at AI Engineer New York 2026. Flows let Claude choose the organization of agents for workloads longer than a single agent or roster.
Latent Space·2 hours ago
Essay urges developers to avoid OpenAI and Anthropic models
Security engineer Micah Lee argues open-weight models now match the latest ChatGPT and Claude on complex tasks like vulnerability research, citing private hosting via secure enclaves (Tinfoil) and provider switching as advantages.
Hacker News·2 hours agoDeveloper builds receipt-printing mystery game on M5StickS3 with Codex
dani bauer·2 hours ago
Perplexity Projects implements Git-based filesystem for agent collaboration
Perplexity·2 hours ago
Microsoft and Musk race to reinvent the AI chip
Bloomberg's Ed Ludlow breaks down Microsoft's latest homegrown Maia chips, aimed at reducing reliance on outside semiconductors for AI. Elon Musk is testing as many as ten new chip designs a year.
Bloomberg Technology·2 hours ago
OpenAI reopens Codex Ambassador applications
OpenAI Developers·2 hours ago
OpenAI GPT-6 Sol and Luna reach Bedrock via cross-Region inference
Amazon Bedrock now serves OpenAI's GPT-6 Sol and GPT-6 Luna through global cross-Region inference for teams in Thailand, Malaysia, Singapore, Indonesia, and Taiwan. The setup targets throttling that occurs when single-Region inference demand exceeds local capacity.
AWS AI Blog·2 hours ago
