|
⏱ The 10-Second Version
- Researchers cracked the encryption protecting OpenAI’s, Anthropic’s, and Google’s AI “reasoning,” then used a cheaper sibling model to decode it, pulling 62 real API keys and 33 real passwords out of public logs.
- Bitcoin miner Riot Platforms signed a 20-year, $9.1 billion compute deal with Anthropic, leasing 191 megawatts of power with 96 megawatts online by the end of 2027.
- OpenAI’s only dedicated ethicist left after less than a year and hasn’t been replaced, with the company saying ethics work is now “deeply embedded” across research teams instead.
|
|
The Big Thing
Researchers Crack the “Private” Reasoning Behind OpenAI, Anthropic, and Google
What Happened
OpenAI, Anthropic, and Google all encrypt a model’s hidden “thinking” before sending it to developers, but researchers found every model in a family shares the same encryption key. That let a cheaper sibling model decrypt a stronger one just by being asked to transcribe it: Claude Haiku 4.5 decoded Opus 4.8, GPT-5.6 Luna decoded GPT-5.6 Sol, Gemini Robotics 1.6 decoded Gemini 1.5 Flash. Scraping 6,708 public agent transcripts, researchers from ELLIS Institute Tübingen, the Max Planck Institute, MATS Research, and Snyk decoded 315,320 hidden reasoning blocks and recovered 704 privacy artifacts nobody meant to publish, including 62 API keys and 33 passwords, sitting only in the “invisible” reasoning step, never the visible reply. All three labs shipped fixes after the report landed. Read the full story at The Hacker News →
Why You Care
Ask your team this week whether anything sensitive has ever passed through an AI agent’s reasoning step, not just its final answer.
|
What to Say in Your Next Meeting
“The ‘private’ reasoning behind OpenAI’s, Anthropic’s, and Google’s answers wasn’t private, and researchers used it to pull real API keys and passwords out of public logs.”
|
|
|
Speed Round
OpenAI expanded its Daybreak cybersecurity program and released GPT-5.6 Cyber, a purpose-built model for vetted defenders doing incident response, malware analysis, and vulnerability research, warning that “threat actors will increasingly use AI to conduct cyberattacks at unprecedented speed and scale”; reported partners include Accenture, IBM, CrowdStrike, and Cloudflare. OpenAI built a model to catch AI attackers the same week researchers cracked its own model’s encryption.
Read the full story at TechCrunch →
Bitcoin miner Riot Platforms signed a 20-year, $9.1 billion compute deal with Anthropic, worth up to $16.1 billion with two five-year extensions, leasing 191 megawatts of data center capacity with 96 megawatts online by the end of 2027; an analyst pegged Riot’s total contracted data center revenue, across this and its existing AMD deal, at $9.8 billion. Riot spent a decade mining bitcoin to end up renting rooms to a chatbot.
Read the full story at CNBC →
Chloé Bakalar, OpenAI’s only dedicated ethicist, left the company in July after less than a year and hasn’t been replaced; OpenAI says ethical considerations are “deeply embedded into the model building process” across research teams rather than owned by one role. OpenAI’s dedicated ethics team is now zero people, the same week its models got easier to read than they were supposed to be.
Read the full story at Gizmodo →
England and Wales’s courts service banned Meta’s camera-equipped smart glasses from every courtroom, citing the long-standing ban on recording proceedings; anyone wearing a pair gets it confiscated at the door and returned on the way out. Turns out the simplest fix for a hidden camera is a sign at the door, no encryption required.
Read the full story at Engadget →
Chinese humanoid robot maker Unitree’s Shanghai IPO was oversubscribed 8,288.82 times by retail investors, pricing the company at roughly $9 billion, 219 times its projected 2025 earnings. Retail investors out-hyped the robots by a factor of eight thousand.
Read the full story at TS2 Tech →
|
|
Tool in 60 Seconds
Gitleaks, the Free Scanner for What Your AI Already Leaked
Researchers just proved AI reasoning logs can hold real API keys and passwords, sitting in text nobody thought to check twice. If your team saves chat transcripts, agent runs, or shared prompts in a repo, doc, or wiki, the same mistake is one paste away. Gitleaks is a free, open-source command-line tool that scans any text for hardcoded secrets in about sixty seconds.
- Install Gitleaks (or run it via Docker) and point it at wherever your team saves AI transcripts: a repo, an exported doc, a shared log dump.
- Review anything it flags before assuming it’s a false positive.
- Paste what it finds into an assistant with the prompt below to prioritize what to rotate first.
Gitleaks just flagged these possible secrets in my team's
AI chat logs:
[paste: e.g. flagged strings, file names, or a summary
of the gitleaks output]
Tell me:
1. Which of these look like real, still-valid secrets worth rotating today
2. What could go wrong if each one is real and already sitting in someone else's hands
3. The one to rotate first
|
|
Forward this to the coworker who pastes the whole AI chat log into a shared doc without a second look.
— Don.
|
