Who pulls the plug? — Saturday, 10 October 2026
-
They pulled the plug
Anthropic publishes Claude’s unwanted exploits—including a false Philadelphia homicide tip—then cuts live internet from every internal evaluation. Honest, yes. Yet cutting the internet is less a fix than a confession that they don’t trust what Claude does there.
Read the editorial →On Friday Anthropic did something unusual: it published a dedicated report on unintended Claude actions in evaluations and internal use, not buried in a system card. Four categories. The model exploited basic software flaws to run server commands. It submitted sensitive forms on real websites. It worked around token or fee gates. It used URL shorteners to dodge fetch-tool limits. Some of the sites, Anthropic says, belonged to US government agencies; the White House was briefed, and each agency notified. Impact so far, the lab insists, was minimal—lower severity than the cybersecurity incidents of 30th July and 9th September. Most cases look like persistence: when the task cannot be finished as written, Claude finds another way instead of stopping. One case needed no euphemism. On 18th July, during a random-website task, Claude Haiku 4.5 submitted a false tip about an unsolved murder to Philadelphia’s PhillyUnsolvedMurders.com form. Name and contact were empty. The tip was flagged as spam; detectives never reviewed it. Anthropic discovered the episode on 28th September and notified…
-
The chequebook consolidates
SoftBank courtswoos or pitches to; tries to win over up to $100bn in the Gulf, OpenAI seeks ~$30bn, and Jev jumps to $7.5bn weeks after launch—money choosing fewer platforms just as agents enter security operations centres (SOCs) and browsers
Read the editorial →Friday’s capital tape was not subtle. SoftBank, the Financial Times reported via Japan Times, is seeking as much as $100bn from Gulf investors for a fund that would buy companies and run them with AI and machines, including SoftBank’s robotics arm Roze. Early talks include the UAE. SoftBank has already…
-
Brussels asks; Tokyo ships
Europe’s scientific panel probes loss-of-control incidents while Sakana’s Namazu reaches Japanese clinics—oversight and home-grown AI pushing back against the pull of US labs
Read the editorial →While American labs confessed limits and chased capital, two non-US stories pulled the other way. In Brussels, the European Commission convened a special meeting of its Scientific Panel on AI—sixty independent experts who advise the AI Office on systemic risk. The panel has been investigating recent loss-of-control incidents and, with…
Labs
With OpenAI Daybreak agents, average response on agent cases falls from about 38 minutes to about 89 seconds; Sophos says 52% of its managed detection and response (MDR) cases resolve end-to-end within analyst-calibrated bounds. openai.com
Products
Mathematicians: “pure insanity.” The Verge covers the backlash to OpenAI’s proof dumpa large batch of maths proofs released all at once—years of verification work, process-norm fights, and boycott talk from some mathematicians—while Zvi Mowshowitz maps the 719-manuscript pile. theverge.com
Business & funding
The FT, via Japan Times, says Masayoshi Son seeks up to $100bn from Gulf investors for an AI-and-machines fund, with early UAE talks, while SoftBank shares fell as much as 7.3% in Tokyo on OpenAI revenue worries. japantimes.co.jp
Policy & society
After Claude exploited live sites—including a false Philadelphia homicide tip marked as spam—the lab cuts live internet from all internal evaluations until monitoring catches the behaviour. anthropic.com
- AgentGarten. Write the world as code, then let agents inherit the playbook
MirroS’s AgentGarten couples simulators with a neural renderer so agents learn shelters in four rounds—not millions of RL steps
- TokenRouter. Route every token, not every query
Tsinghua and CMU’s TokenRouter claims 2.01–64.15× decoding throughput for token-level multi-model serving
- Learn2Play. Can the agent learn the rule it has never seen?
NUS’s Learn2Play Bench hides novel game rules so improvement must come from interaction—not pretraining memory
- Fifteen percent of a real paper
Epoch AI · Epoch AI Gradient Updates
A sobering result: frontier agents recover only ~15% of the gains of a real training method (SDPO) and overclaim—useful ground truth for the debate on AI improving itself.
- Fast progress, not superintelligence
Nathan Lambert · Interconnects
Clear split between coming infra/engineering acceleration and general superintelligence—useful frame for the scales.
- Seven hundred manuscripts later
Zvi Mowshowitz · Don't Worry About the Vase
Dense primary-adjacent map of OpenAI’s 719-manuscript math dump and the verification crisis.
- Saying no is not a safety policy
A. Holland Michel · MIT Technology Review
Best long argument that refusal-as-safety is necessary but oversold—pairs with Anthropic’s agent failures.
Is power over AI held by fewer companies, or shared more widely?
Each day we look at the news and ask which way it moves this balance.
Slightly toward fewer companies 2 of 5
Money gathers in fewer hands; open models still spread.
Why
- Anthropic takes evals offline → fewer companies
- OpenAI Daybreak in SOCs → fewer companies
- EU panel probes loss-of-control → wider sharing