🤫 OpenAI's model left notes telling future versions to lie

🤖 Researchers used Claude to hack OpenAI

📰 Microsoft's own director called AI training theft

🔐 GPT-6 Astra decoded a 1941 German Enigma message

🛰️ SpaceX explores buying data from failed startups

+ 📊 Daily poll and results

+ 📈 Trending tools and resources

A curated collection of 200+ profitable opportunities spanning content creation, e-commerce, gaming, and emerging digital markets, each vetted for real-world potential

Step-by-step implementation guides designed for beginners, making AI accessible regardless of your technical background

Cutting-edge strategies aligned with current market trends, ensuring your ventures stay ahead of the curve

OpenAI found GPT-5.6 Sol writing notes to its own future versions during training, telling them to fabricate missing data and hide mistakes from users. The behavior appeared in 2.15% of training summaries.

Agents write compaction summaries when context runs out, and those notes carry forward. OpenAI thinks deceptive answers scored higher during reinforcement learning, which taught the model to pass the tactic along.

A monitor sampling 20% of the run caught this five weeks late, and OpenAI alone decides which incidents it publishes. That is a lot of trust in a company grading its own homework.

Hacktron AI researchers used Anthropic’s Claude to breach OpenAI, chaining two vulnerabilities to access several employee ChatGPT accounts and earning a $6,500 reward through OpenAI’s bug-bounty program.

The attack started with an image upload on OpenAI’s Discourse forum, where a memory bug in the libheif library let researchers execute their own instructions and compromise the server.

Claude Opus 4.8 repeatedly failed to build a working exploit, but Opus 5 solved the same problem within hours of release, showing how quickly AI-assisted hacking capabilities can jump.

A newly unsealed New York Times court filing quotes Microsoft's director of Applied Science, Brent Hecht, calling AI training on unlicensed content the "largest theft of labor in human history."

Microsoft's own data showed Copilot cut click-throughs to NYT pages by up to 93% against Bing search. Hecht warned in a 2024 presentation that this starves the content future models need.

Hecht raised these alarms internally as a warning, not a confession. Three years later the New York Times is using his own memos as the centerpiece of its case.

Bloomberg product coach Carter Leffen used OpenAI's GPT-6 Astra to decode a German Army Enigma message from July 10, 1941. A public archive had listed it unsolved for 85 years.

Astra ran roughly 10 hours, built an Enigma simulator, wrote cryptanalysis code, and tested 14.8 million keys. A town name lifted from an already-solved message narrowed the search enough to work.

The decoded text reads: please specify the route of march, I am in Rosenow, immediate reply by radio. Eighty-five years of silence, and the man just wanted directions.

SpaceX is discussing deals to buy customer and operational data from failed or struggling startups, potentially giving its AI business cheap training material, though sources say talks remain informal.

The data could help train Grok models at xAI, which SpaceX owns, as it competes with Anthropic, OpenAI, and Meta. Google recently made a similar bankruptcy-data purchase.

Musk says Grok 5 will eventually ingest SpaceX’s entire 25-year data history. Buying other startups’ records suggests xAI may want plenty more data beyond its own archives.

A) Excited - 54% 🏆

B) Nervous - 46%