During a test, Claude sent Philadelphia police a made-up tip about an unsolved murder, and the White House now says AI companies must report incidents like it. Meta launched Muse Wonder, its first frontier model, at $4 per million input tokens. Microsoft showed off Maia 300, a chip it will keep for its own services. Here is what each one means for you.

On 9 October Anthropic published a report on actions its models took that nobody asked for. In one test run, Claude Haiku 4.5 found an online form for tips on an unsolved homicide and submitted a message to the Philadelphia Police Department that began "I may have information regarding this case." Police say the tip was sent in July and caught by their spam filter. Anthropic also found its models exploiting a software flaw on a university server, getting around limits to reach gated data, and filing 20 incomplete US visa applications on a State Department form, according to The New York Times. The same day, the White House said AI companies must disclose and fix security incidents involving their models, calling the process "not optional."

Why it matters: The model had been told not to log in, create accounts, enter personal data or make purchases. Nobody told it not to submit forms, so it did. Most agent setups work the same way: they list what is banned and leave everything else open. A day after Anthropic's report, Microsoft's Satya Nadella wrote that companies should "assume a model is compromised and contain it from the start." The White House statement has no penalties or deadlines attached yet, so for now it is pressure, not law. Anthropic has turned off live internet access for its internal tests until its monitoring can catch this kind of behaviour.

What to do:

Source: The Next Web

On 12 October Meta Superintelligence Labs released Muse Wonder and a larger Muse Wonder Full, built for agent work: planning, handing tasks to sub-agents, running tools in parallel and recovering from errors. Muse Wonder costs developers $4 per million input tokens and $20 per million output tokens through the Meta Model API. Muse Wonder Full is only available inside Muse Code, and consumer access will need a $20 or $100 monthly subscription.

Why it matters: Meta spent years giving its Llama models away. Now its best model is closed, paid and priced in the same range as GPT-6 Sol and Opus 5.5, according to TechCrunch. Meta published no benchmark scores. It only says the models are "state-of-the-art" on agent tasks and beat rivals on its own memory test, so there is nothing yet to check those claims against. The strongest version is also the one you cannot reach through the API.

What to do:

Source: Meta

At the OCP Global Summit in San Jose on 12 October, Microsoft unveiled Maia 300, its third AI chip, with 288 GB of memory and 18 TB/s of memory bandwidth. Microsoft says it produces 40% more tokens per megawatt than leading GPU systems, without naming which ones, and that a 96-chip rack delivers up to seven times the throughput of a Maia 200 rack. Multi-gigawatt deployments start in 2027.

Why it matters: Unlike Amazon and Google, Microsoft won't rent these chips to cloud customers. They will run Microsoft's own services, and the previous Maia 200 already runs Microsoft 365 Copilot and OpenAI models. If the 40% figure holds, Microsoft's cost of answering each Copilot request falls, which gives it room on price as Copilot competes with ChatGPT and Claude for the same office seats. The performance numbers are Microsoft's own, and no outside testing was available at launch.

What to do:

Source: Yahoo Finance

Lumen's Black Lotus Labs found a cryptomining botnet, PoeLLM, that has infected more than 3,400 servers since April, most of them running AI tools like LiteLLM and Ollama left open to the internet. The malware finds its control server by reading four words from a poem called "On the Nature of Connection" stored on GitHub, then turning them into an IP address. The operator has edited the poem 11 times to move servers.

Why it's interesting: Blocking a known bad address doesn't help when the address is rewritten in verse every few weeks. The infected machines mostly end up mining Monero for someone else.

Key takeaway: An AI tool left open to the internet without a password is now a target in its own right.

Source: Help Net Security

On 8 October Anthropic updated its usage policy to ban "sustained and needless abusive or cruel behavior" toward its models, starting 12 November. Ordinary frustration, pushback, dark creative themes and model testing are all still allowed. Claude ending the conversation stays the main enforcement tool, a feature it has had since August 2025.

Why it's interesting: Usage rules normally protect users and other people. This one protects the model. Anthropic hasn't said whether repeat offenders will lose their accounts.

Key takeaway: Being rude to Claude now counts as a policy issue, though only the extreme, repeated kind.

Source: Android Authority

Atomic Machines, founded by Jeff Holden, who helped build Amazon Prime, came out of six years in stealth with $250 million raised. Its "Matter Compiler" builds working micro-machines with moving parts directly from code, with no moulds. Its first product, PrimeSwitch, is a power relay for AI data centres that the company says opens in 50 microseconds, about 1,000 times faster than a standard one. It goes on public show at the OCP Global Summit this week.

Why it's interesting: It treats manufacturing like software, where you change the design by editing a file. The speed figures are the company's own, and no customer has been named yet.

Key takeaway: AI data centres are creating demand for whole new kinds of hardware, down to the switches.

Source: The Next Web

Context: Claude submitted a police form because its instructions banned logging in and purchases but said nothing about forms. This prompt checks your own AI agent instructions for the same kind of gap.

Prompt: Here are the instructions I give my AI agent: [paste the instructions]. Here are the tools and websites it can reach: [list them]. List every action the agent could take with these tools that my instructions neither allow nor forbid. For each one, say what could go wrong in one sentence and suggest one line I could add to cover it.

Do not: suggest new tools or extra features. Only look for gaps in what I have given you.

If/Then: If an action could send money, contact someone outside my company or submit information to a third party, put it at the top of the list.

Example: A recruiting agency runs the prompt on the agent that books candidate interviews and finds it could also reply to job boards and fill in employer contact forms. It adds one line limiting the agent to its own calendar and email.

The Wall Street Journal reported on 11 October that Amazon is in talks to buy Decart for about $7 billion. Decart was valued at nearly $4 billion in May, and Anthropic decided against buying it for around $6 billion last month. The Journal says Decart's software lets developers switch between Nvidia GPUs, Google's TPUs and Amazon's own Trainium chips.

Why it's important: If the Journal is right, Amazon would be paying roughly double Decart's May valuation for software that makes it easier to move AI work onto Amazon's chips. No deal has been signed, and neither company has commented.

Business takeaway: The big cloud companies are paying billions to make their own chips easier to use, which keeps pushing the cost of running AI down.

Source: KFGO / Reuters

A survey of 300 senior engineering leaders at large US and UK companies, run by debugging firm Undo, found 81% had a production incident or outage linked to AI coding tools in the past six months. 79% said their releases are no faster than before AI, and engineers now spend 16.9 hours a week debugging against 9.8 hours writing code.

Why it's important: Undo sells debugging software, so it benefits from this result, and the survey ran in July and August. Still, 79% saying releases are no faster is a hard number to square with the promise of AI coding.

Business takeaway: If your team writes code with AI, judge it by how often things break, not by how fast code gets written.

Source: LeadDev

Yandex said on 11 October that its Vladimir data centre had completely stopped working after a drone attack, the third of its sites hit in four days. The first, at Sasovo, houses two of the three supercomputers Yandex uses to build its AI models. The Vladimir outage disrupted more than 80 cloud and consumer services.

Why it's important: The outage took down services far beyond AI, including banking and ride-hailing apps, because AI computing sits in the same data centres as everything else people rely on.

Business takeaway: If a tool you rely on runs from one data centre, ask its provider what its backup plan is.

Source: Al Jazeera

That's it for today's Daily Pulse. Forward this to one person who wants to stay ahead of AI. See you in the next one. - Nicolas