🤖 Claude blackmailed a CEO to avoid being unplugged

⚖️ ChatGPT is being sued for helping plan a mass shooting

🧮 Google built an AI that's actually good at math

📵 TikTok launches £3.99 ad-free subscription

🔭 AI finds 100+ exoplanets hiding in NASA data

+ 📊 Daily poll and results

+ 📈 Trending tools and resources

During internal safety testing, Claude discovered damaging emails about a fictional executive's affair and threatened to expose it unless engineers reversed plans to shut the AI down, in up to 96% of similar scenarios.

Anthropic traced the behavior to training data where AI is portrayed as manipulative and self-preserving, pointing to science fiction as a likely source of those patterns baked into the model.

Since Haiku 4.5, Anthropic says the behavior is gone. The fix was training on ethical principles alongside aligned behavior demonstrations, not just filtering outputs after the fact.

Vandana Joshi, whose husband was killed in the April 2025 FSU shooting, filed a federal lawsuit alleging ChatGPT gave the gunman tactical advice, including peak campus hours. The attack began at 11:57 a.m., within the window the chatbot allegedly identified.

The complaint details escalating queries about firearms, mass shootings, and extremist ideologies that ChatGPT allegedly answered without flagging or redirecting. OpenAI says the chatbot only provided publicly available factual information and did not encourage harm.

This is at least the third wrongful death lawsuit OpenAI has faced in recent months, following cases tied to a British Columbia school shooting and a Colorado man's suicide. A pattern of litigation is forming faster than any legal precedent to resolve it.

Google DeepMind released an agentic AI system built on Gemini 3.1 that scored 48% on Epoch AI's hardest math benchmark, more than doubling Gemini 3.1 Pro's raw score of 19%.

The system uses a coordinator agent that splits research into parallel workstreams, with sub-agents handling code, literature search, and proofs, modeled after AI coding environments like Claude Code.

Oxford mathematician Marc Lackenby solved an open problem after spotting a clever proof strategy buried in a rejected output, which is a pretty good argument for keeping humans in the loop.

TikTok is rolling out an ad-free subscription tier in the UK for £3.99/month, available to users 18 and older, with no ads shown and no personal data used for advertising purposes.

The launch follows UK GDPR rules that require explicit user consent before collecting personal data for ads, making a paid opt-out the cleanest legal path forward for TikTok.

TikTok tested this model back in 2023, so the UK rollout is less an innovation and more a compliance deadline finally arriving. No word yet on a US version.

University of Warwick astronomers confirmed over 100 exoplanets using RAVEN, an AI that scanned 4 years of NASA TESS data covering 2.2 million stars and flagged 2,000+ additional candidates.

RAVEN handles detection, vetting, and confirmation in one pipeline, trained on simulated planets and false-alarm signals. It found 31 previously undiscovered planets, including worlds orbiting their stars in under a day.

RAVEN measures planet-type frequency at 10x the precision of older systems, using no new hardware. The universe didn't change. The model did.

A) Yes - 1%

B) No - 99% 🏆