A New Attack Just Cracked Every Major AI Lab’s “Encrypted” Chain of Thought — And Leaked Passwords Doing It
A wild week closes with a security bombshell, two pricing wars, a profitability milestone, and a billion-user product moment. Here’s everything that moved in AI over the past few days.
Researchers Crack “Encrypted” Reasoning Across Anthropic, OpenAI, and Google
A new paper shows that provider-issued encrypted reasoning blocks — the “chain of thought” text labs hide from users to protect their models — are interchangeable across sessions, users, and models within the same ecosystem. That means an attacker can inject a more capable model’s encrypted reasoning into a weaker sibling model and force it to decrypt in plaintext, without ever jailbreaking the capable model directly.
315,000 Reasoning Blocks Decoded, 182 Credentials Recovered
The researchers demonstrated the trick across all three major labs, decoding 315,320 encrypted reasoning blocks pulled from public repositories. In the process they recovered 367 pieces of personally identifiable information and 182 credentials, and identified a route for invisible prompt injections that can persist inside agentic workflows. It’s a serious wake-up call for anyone treating “encrypted reasoning” as a hard security boundary rather than an obfuscation layer.
Anthropic Posts First Operating Profit — Two Years Ahead of Schedule
Anthropic’s Q2 2026 revenue hit $10.9 billion, up 130% year over year, and the company posted its first operating profit of $559 million — a milestone it wasn’t projecting to hit for another two years. The caveat: a SpaceX compute-ramp discount likely flattered the quarter, and analysts are already flagging Q3 margin pressure once that discount rolls off.
ChatGPT’s Free Tier Just Got a Lot More Generous
OpenAI made GPT-5.6 Luna the new default model for ChatGPT’s free tier, following an 80% price cut on the underlying API rate (from $1 to $0.20 per million tokens) that took effect July 30. Free users now get unlimited chats on Luna rather than a capped daily quota — a direct shot at Gemini and DeepSeek’s push into the free-tier market this month.
DeepSeek Raises API Prices 93% — Still Far Cheaper Than Rivals
In the opposite direction, DeepSeek raised pricing on its V4 Flash model by 93%, from $0.14 to $0.27 per million tokens, effective August 14. Even after the hike, DeepSeek remains dramatically cheaper than frontier-lab pricing, and the move is being read as a sign that DeepSeek’s earlier prices were unsustainably low rather than a retreat from the low-cost strategy.
Gemini Crosses 1 Billion Monthly Users, Catching ChatGPT
Sundar Pichai announced on August 11 that the Gemini app passed 1 billion monthly active users, calling it Google’s fastest-growing product ever and the 14th Google product to hit that threshold. TechCrunch reports 63% of users now engage voice features and Gemini is generating 150 million-plus images daily — putting it roughly on pace with ChatGPT, which crossed the same milestone back in June.
That’s today’s wrap. The encrypted-reasoning attack is the one worth watching closest — it’s the first serious crack in a security assumption every major lab has been leaning on, and the fallout (patches, disclosures, possibly new encryption schemes) will likely dominate coverage for weeks. Follow NeuralPaws daily for the next AI news drop.