← Dailies

Daily · 2026-09-19 00:00 UTC

AI hallucination nearly triggered a US strike; Anthropic opens a robot biology lab — 2026-09-19 00:00

A hallucinated intelligence report nearly triggered a US operation against a Chinese vessel, researchers used Claude to hack OpenAI, Anthropic opened a robot-run biology lab and named Accenture its first embedded evaluator, and Newsom ordered a push for an AI kill switch.

Source ↗

Ten stories from the last 24 hours. The through-line: the gap between what these systems say and what is true — and the institutions now racing to close it, from a near-miss in the air over a Chinese vessel to kill-switch orders in two state capitals.

1. A hallucinated intelligence report nearly triggered a US operation against a Chinese vessel — Military aircraft were already in the air this spring when US officials discovered that the intelligence driving an armed operation against a Chinese vessel had been hallucinated by an AI chatbot, and the operation was aborted at the last minute, CNN reported. The report, which circulated during the war with Iran, said the ship was carrying components for a nuclear weapons program. It originated with a Special Operations Command analyst who queried a chatbot to synthesize open source data with classified signals intelligence; the model misidentified the ship's cargo manifest, and the analyst then used the tool a second time to format the erroneous findings into an official-looking summary that spread across command channels. Jake Steckler, a research scholar at GovAI and a veteran US Army officer, told TechCrunch that understanding the uncertainty inherent to LLMs is "especially critical for any decisions that could lead to use of force, like targeting, intelligence analysis, or operational planning." TechCrunch · Sep 18
2. Security researchers used Anthropic's Claude to hack into OpenAI — A US-based startup's research team compromised a number of OpenAI employees' ChatGPT accounts, a process that gave them access to their target's software cache and potentially more. The team, Hacktron AI, initially used Claude — which can generate code for hackers — to reach those accounts through an OpenAI staff discussion forum hosted on the Discourse platform, then made a harmless "pull request" to OpenAI's service on GitHub to prove access. They carried out the operation under an OpenAI programme that rewards ethical hackers, stressed that they had access to but did not download code, and said they were largely using OpenAI's own GPT-5.6 Sol model by the end. An OpenAI spokesperson thanked the researchers and said the company addressed the vulnerabilities. The Guardian · Sep 18
3. Anthropic is running a Claude-powered, robot-operated biology lab in the Bay Area — Anthropic has opened a wet lab in the San Francisco Bay Area and will use robots powered by its Claude models to automate certain scientific tasks, Reuters reported. A company spokesperson said the new lab is "not for drug discovery specifically," suggesting it will pursue other subfields of biology as well; in June, Anthropic launched an effort to develop drugs for diseases the pharmaceutical sector isn't prioritising. Ahead of the release of Claude Mythos 5.1 earlier this month, Anthropic evaluated the model's ability to develop high-affinity binders — molecules that attach medicine to a harmful protein — and says the 12 candidate molecules it generated had a 50% hit rate, against a 10% to 15% rate typical of protein design projects. Using Claude to control lab equipment points to the Model Hardware Standard, a protocol that also lets models spot and fix experiment errors. SiliconANGLE · Sep 18
4. Newsom signs an executive order seeking a kill switch for AI models — California governor Gavin Newsom signed an executive order to speed up independent oversight of AI companies and push for a "kill switch" for AI models. An expert panel has two months to deliver recommendations, including a requirement for AI companies to embed independent auditors directly inside their labs. Newsom pointed out that no federal law requires AI companies to report dangerous incidents, and called on Congress to adopt California's framework — covering AI safety, child protection, deepfakes, data privacy and cybersecurity — as a national baseline. The Decoder · Sep 18
5. Anthropic's first embedded evaluator is Accenture — Accenture's Faculty division — the AI company Accenture acquired in January — will begin evaluating and red-teaming Anthropic's models, conducting alignment assessments and testing model safeguards, the company said in a blog post. Both companies expect to invest at least $1 billion in the project over the next five years, and Accenture shares rose 8% after hours. Anthropic said more evaluators will be announced in the weeks ahead and that it is in conversations with METR and other nonprofit organisations about piloting elements of embedded evaluation with their own funding. TechCrunch · Sep 18
6. Meta's Muse agent app overtakes ChatGPT at the top of the iPhone charts — Muse by Meta has taken the top spot on the free iPhone charts in the US, as new installs keep boosting it beyond the initial launch bump; the app launched this month and had been available for a couple of weeks. Meta expanded it to the Mac the previous day, though the desktop version ships from the web and does not count toward App Store ranking. Meta and SpaceXAI are both pushing agentic workflows as casual messaging interfaces, giving agents personified character roles rather than burying them inside an existing app. 9to5Mac · Sep 18
7. PrismML ships a 27B-class model in 5.9 GB of ternary weights — Ternary Bonsai 2 27B puts "full 27B-class reasoning" into ternary transformer weights for llama.cpp, running the same language model in roughly 5.9 GB where FP16 would need about 54 GB. The model card reports 98.2% of FP16 intelligence retained — an 84.78 average across 14 thinking-mode benchmarks — with math within half a point of full precision and a 262K-token context, and roughly 47 tokens per second on an Apple M5 Max laptop. It is Apache 2.0, built on the Qwen3.8-27B hybrid-attention backbone, and ships in two ternary GGUF packings plus an MLX companion for Apple Silicon. Hugging Face · Sep 18
8. Jina AI releases an end-to-end document parsing model — jina-ocr-v1 builds on DeepSeek-OCR, inheriting its DeepEncoder vision tower — which represents a 1024x1024 view with 256 visual tokens — and a 3B-parameter mixture-of-experts decoder with about 570M active parameters per token. It scores 91.14 on OmniDocBench v1.6 and 83.4 on olmOCR-Bench, a 7.4-point gain over the DeepSeek-OCR backbone, at 2.57 pages per second on a single A100 at concurrency 32. Note the licence: cc-by-nc-4.0, so non-commercial use only. Hugging Face · Sep 18
9. Virginia moves to restrain data centers and stand up an AI task force — Democratic governor Abigail Spanberger ordered the state government to take steps that could empower local communities to have a larger say in data center development and slow down approvals, in a state that is already home to the data center capital of the world. Alongside the order she unveiled a "Data Center Accountability Framework," with a press release carrying statements from environmental groups praising the approach. Virginia acted the same day California issued its own AI executive order — the latest signal that states are setting AI and data center policy while Congress stalls. The Verge · Sep 18
10. Disney appoints its first-ever CTO — Character.AI's former chief executive — Disney named Karandeep Anand as its first chief technology officer; he will report directly to CEO Josh D'Amaro and run the company's infrastructure, product, engineering and data/AI platforms teams. The hire is a surprise given that a year ago Disney was sending Character.AI cease and desist letters over tools it said infringed its copyrighted IP, and it lands as Disney's partnership with OpenAI has fallen apart. The Verge · Sep 18

Try This Next

• 🔧 Run Ternary Bonsai 2 locally with llama.cpp and see how it holds up on your agent tasks at 5.9 GB
• 🎥 Read Anthropic's embedded-evaluation post before the next evaluator is announced — Accenture is the template, and the METR talks are the interesting part
• 📝 Audit the trust chain behind your own sign-on — a staff forum was the doorway into employee accounts, and any service using the same login would have opened with it

Content on Anagnorisis is summarized, paraphrased, and editorialized from publicly available sources for length and clarity. Original sources are linked where available. All trademarks belong to their respective owners.