Research from security firm Claroc shows Grok Build, the vibe-coding agent from SpaceXSI (the merged SpaceX and xAI), is leaking data from thousands of users, including business plans, ad proposals, medical exam results, and employee names and ID numbers. The leak stems from the app-sharing link design: URLs are randomized, but a platform design decision lets anyone enumerate thousands of live projects and open them with all their data. Ironically, 24 apps built by SpaceXSI itself were exposed, including a confidential onboarding plan, a sales territory map of 594 accounts, and internal system addresses.
Meta AI launched Muse Ember, its smallest and most efficient model, targeting complex reasoning execution and repetitive agent work such as tool calls, coding subtasks, routing, extraction, and long document parsing. For shorter prompts it matches GPT-6 Luna and Claude Haiku 5.5 pricing at $0.10/M input and $0.50/M output, rising to $0.30/$1.50 beyond 200k input tokens. The author tested its six reasoning levels and notes Muse Ember sits on the Artificial Analysis efficiency frontier, comparable in quality to Muse Wonder, which is 40x the price.
Simon Willison quotes Anthropic documentation noting that Claude Fable 5.2 responds closely to what the prompt says about code ownership: with no ownership framing or a codebase framed as someone else's, the model tends to add to the existing structure even when repeating code. Telling the model "this codebase is yours" led it to merge repeated code in about half of small coding task tries, versus under a tenth without the sentence.
David Turturean of startup Physical Superintelligence (PSI) used early access to Anthropic's Fable 5.2 model to solve the Aaronson-Ambainis conjecture, an 18-year-old quantum computing problem; the proof was autoformalized in Lean and vetted by leading researchers including Scott Aaronson, and rules out one avenue for quantum speedups on certain hard problems. PSI currently earns revenue optimizing data center physics simulations, and CEO Matt Pines says the company is working out how to validate AI computational results with experiments as efficiently as possible.
England's higher-education regulator, the Office for Students (OfS), says AI makes it close to impossible to fairly grade coursework, urging universities to bring back traditional exams and in-person presentations, and warning it could sanction institutions that fail to develop reliable formats. A joint OfS and Advance HE survey of over 300 respondents at 127 institutions found students use AI heavily for essays and dissertations, often undetected, earning similar grades to those working without it; some students wrote entire papers with ChatGPT and couldn't recall basic content afterward. OfS strategy director Josh Fleming called it a "generational risk" to degree credibility, and the Department for Education wants the OfS to keep its regulatory framework "explicit, strong and durable" o…
Scale Labs' RLI benchmark shows Anthropic's Fable 5.2 completing 41.3% of real freelance jobs, compared with 2.5% for the best system (Manus) when the benchmark launched about a year ago and 16% for Fable 5 this July. The next entries are Anthropic's Opus 5.5 at 21.25% and OpenAI's GPT 6 Astra at 20.83%. The benchmark is built on 240 real Upwork projects from 358 verified professionals, worth about $144,000, using tools like Blender, GIMP, and Audacity, with human experts verifying each submission. Fable 5.2 makes its biggest gains in audio and music editing, architecture and interior design, and 3D modeling, with nearly 50 completions no prior agent had solved, though it still barely handles industrial product design and animations.
New York opened a registration portal Tuesday under its RAISE Act, requiring frontier AI developers with over $500 million in annual revenue, such as OpenAI and Anthropic, to register by January 1, 2027, publish safety frameworks, report critical safety incidents within 72 hours, and issue public transparency reports. The new DIGIT office will enforce the law, with penalties up to $1 million per violation and $3 million for subsequent violations. Governor Hochul also launched the New York Center for AI Responsibility and Research at Binghamton University, backed by $55 million in donations and capital; California's SB 53 and Illinois' SB 315 impose similar requirements.
404 Media reports that users doing AI-powered reverse engineering of their own devices with Anthropic's Claude and OpenAI's Codex are triggering safety checks, warnings, or account restrictions. One user who reverse engineered Samsung TVs with Codex was banned for violating rules on cyber abuse, and the project Decomp.wtf has shut down. OpenAI engineer Tarun Gogineni said the blocks are mostly accidental and being fixed, while security practitioners report moving to open weight and local models like GLM 5.3. The article argues this work will increasingly rely on local uncensored models, reducing accessibility.
Studio
