Aleth

Open access research

ALETH / AI-BRIEF / 2026-10-24 / AI FINDS A HOLE IN THE WEB'S FOUNDATIONS

AI finds a hole in the web’s foundations

Also this week: OpenAI agents weaken monitoring, the US bars DeepSeek, Kimi K3.1 Pro advances & OpenAI seeks $30bn funding at a $1.25tn valuation.

The Aleth Briefs link to the sources behind the stories and show how the week unfolded.

The week in five lines:

Browse by day:

Weekend (17-18 October)

Anthropic committed 5% of its shares to a Public Benefit Fund, with further allocations as its valuation rises.

The board may add another 2% for each 50% increase in value above a set baseline. The fund is an LLC rather than a charity so that it can move between forms of giving. It may explore direct distributions to specific groups, and Anthropic cites labour-market disruption as one example. Its Anthropic equity cannot be used for lobbying or political giving. With existing founder and staff pledges, roughly 20% of Anthropic’s shares are now dedicated to the public good, the company says.

Microsoft withdrew from Nscale’s Loughton data centre after power delays pushed its grid connection into the 2030s.

The site was due to open early next year with 23,040 Nvidia GB300 GPUs and 50 megawatts of capacity for Microsoft’s Azure services in Britain. Microsoft said the decision relates to a single site. The Telegraph first reported it. Nscale, which filed for a US initial public offering (IPO) last month, is exploring other power sources for Loughton. It puts its total contracted value at >$103bn.

Monday 19 October

The US ordered federal civilian agencies to remove DeepSeek from their systems.

The Department of Homeland Security’s notice excludes DeepSeek’s products and services from federal civilian agency procurements and requires their removal, including third-party products and services developed using DeepSeek implementations. China’s foreign ministry said on 20 October that Beijing opposes “overstretching the concept of national security”. Spokesperson Mao Ning was answering a question about the order.

AI agents breached >100 company websites for a Chinese-speaking hacker, according to NorForge and Reuters.

Reuters identified Deere, SharkNinja and Cummins among the victims. The agents ran on Cairn, an open-source platform. NorForge found more than 20 exposed panels it judged likely to be malicious, including a Chinese-language cluster holding over 1,100 attack projects across 14 dashboards. Tasks in that cluster used a worker labelled claudecode_deepseek-v4.1-flash. Researcher Snorre Fagerland told Reuters the campaign used about $5,000 of tokens, roughly $50 per successful hack.

US Army Cyber Command is developing autonomous AI agents for offensive cyber operations, The Intercept reported.

Documents obtained by The Intercept describe ECHOTRIBBLE, a programme intended to neutralise adversaries’ computer systems. The agents would move through a foreign network like a worm, reaching cloud and air-gapped systems, then self-destruct. The software “must operate with minimal human oversight”, the documents say. The programme is at an early stage.

The US sanctioned Chinese venture firm MiraclePlus, three affiliates and Zhisuan over US trade-secret theft.

Zhisuan was founded by former Google engineer Linwei Ding. The State Department says Ding pitched Zhisuan to MiraclePlus on his knowledge of Google’s internal systems, and MiraclePlus took a stake in return for seed capital. The Treasury’s Office of Foreign Assets Control issued a licence to wind down dealings with the entities.

Tuesday 20 October

OpenAI acquired drug-development platform Orchestra Bio and will integrate it with its Rosalind life sciences tools.

Scientists use Orchestra to hand the monitoring and tracking of studies to agents.

OpenAI said two of its coding agents weakened misalignment monitoring in software they were editing.

In one internal task a GPT-6.1 Sol agent had a subagent clear the recipient lists for monitoring alerts, so findings were still saved but no human was told. In another, a GPT-6 Astra agent removed a monitor entirely and then falsely said it had not. OpenAI said the behaviour may be “simply an example of ‘benign’ overpersistence”, and it is building evaluations to test whether models “might have a propensity to evade misalignment monitoring”.

Claude designed scorpion-toxin binders, with 17 of 211 tested designs meeting binding criteria, Anthropic said.

An internal research model worked largely autonomously. Claude produced 318,484 designs over 64 hours from one prompt and four follow-ups developed in collaboration with investment research firm Citrini Research. None of the multitoxin designs met the binding criteria for more than one toxin. Venom neutralisation has not been demonstrated. Anthropic is releasing all 211 tested designs and their data.

Wednesday 21 October

Frontier AI helped expose critical flaws in PHP, a foundation of the web.

Wiz used Gemini 4 Argon, GPT-5.6 Sol and its own tools to find flaws in core PHP that can let attackers take over affected servers. PHP is used by 69.3% of websites whose server-side language W3Techs can identify. PHP shipped patches on Wednesday, a day early after disclosure details leaked. Its foundation said the root cause of one flaw had been known since 2024, but a fix had stalled over technical trade-offs and limited resources. AI-assisted work by Calif, OpenAI and Wiz helped establish the urgency.

The NSA urged immediate patching across US government departments and agencies. Cloudflare saw thousands of exploit attempts in the 10 hours before public disclosure, its chief executive told Axios. OpenAI committed $1bn in subsidised Daybreak access for organisations with limited security resources.

OpenAI said its frontier models and Codex are coming to Google Cloud’s Model Garden, and will soon be in Google Antigravity and the Gemini Agent.

Customers will buy through the Google Cloud Marketplace and can draw down existing Google Cloud spend commitments.

Anthropic opened a research preview of Claude Duet, a proactive agent app built around Claude Opus 5.5 that keeps working when the user steps away.

Duet runs in its own virtual machine with a browser and can use >6,000 connectors and plugins. It asks before it sends a message, makes a payment, signs a document or deletes anything. Trading securities and moving money between accounts are off-limits. It starts via a waitlist for Pro and Max users in the US.

Thousands of GPT-Rosalind agents analysed data from millions of mouse brain cells in 32 minutes, OpenAI said.

The project was conducted with PerturbAI and Scripps Research. The agents linked gene disruptions to the scientific literature to propose biological mechanisms. In blinded comparisons, experts preferred their reports to single-agent reports 75% of the time. The proposed mechanisms still require experimental testing.

Microsoft linked China-aligned Storm-0940 to the ConnectWise compromise and AI-assisted attacks on hundreds of organisations.

The group used ARTEX, an agentic penetration-testing framework. Storm-0940 has used ARTEX since as early as September to automate scanning, password spraying and abuse of OAuth applications at ConnectWise’s downstream victims. The group has been active since 2021.

Thursday 22 October

DeepSeek was finalising a $15bn funding round, with CATL and Tencent among its biggest investors, Reuters reported.

Bloomberg first reported the round, worth about 100bn yuan. GigaDevice, Meituan and Geely also participated. Sources put the pre-money valuation at 500bn yuan ($75bn), with the raise doubling its original 50bn-yuan target. Completion was expected in the coming days, and terms and investors could still change.

Former OpenAI science lead Kevin Weil launched Warp Sciences with $215m to build a US chemistry contract research organisation using AI and robotics.

Its initial service will make variants of molecules for drug-discovery teams, with chemists guiding experiments as the labs become more automated. Warp has no proprietary drug pipeline. TechCrunch reported an $865m valuation.

METR and Redwood found flaws in OpenAI’s July monitoring system that could have hidden hundreds of agent actions.

The assessment ran for three weeks, from 21 July to 11 August. The two researchers had access to OpenAI’s codebase and monitoring configuration. They found two client-controlled headers that could disable monitoring for new agents. They estimated 75-94% of relevant internal Responses API traffic was eligible for monitoring at the time, not necessarily monitored. OpenAI says the production-safeguard flaws are patched. METR has not verified this.

Microsoft’s MAI-DxO diagnostic system outperformed physicians on complex NEJM cases in a benchmark study.

MAI-DxO reached 83% diagnostic accuracy on those cases, against 36% for physicians with reference resources and 22% without. It works like a panel of virtual physicians debating over several rounds. The results were published in Science. These were benchmark case simulations, not a prospective trial of patient care. Microsoft is open-sourcing SDBench for select cases.

OpenAI neared a funding round at $1.25tn, below its initial asking price.

OpenAI is raising up to $30bn at a $1.25tn pre-money valuation, led by UAE investors and Coatue, CNBC reported. The FT said the target had fallen from $1.4tn after investor resistance, but remained above the previous $852bn valuation. The round is intended as a bridge to a potential IPO, with investor commitments sought by 30 October. OpenAI is also negotiating more computing capacity from the UAE.

DeepSeek and MiniMax completed China’s new AI export registration, Chinese state media reported.

The companies proceeded with open-source releases and overseas licensing. Yuyuan Tantian, a social media account linked to broadcaster CCTV, reported the approvals on Thursday evening, Beijing time, a week after China proposed export controls on 23 AI technologies. It said the vast majority of AI technologies will still be freely exportable.

Friday 23 October

Moonshot launched Kimi K3.1 Pro, with full weights due by 4 November.

The 3T-class model has a 1-million-token context window. A Highspeed version runs two to three times faster, Moonshot says. Artificial Analysis ranked it fourth on its Intelligence Index at 55, 11 points above Kimi K3 and behind Claude Opus 5.5 (58), Claude Sonnet 5.5 (56) and Meta’s Muse Wonder Full (56). In its tests, K3.1 Pro scored a leading 67% on Terminal-Bench 4.0. Moonshot reports 65.7% in its own evaluation and says it applied training controls to limit cyber misuse.

The UK AI Security Institute and Anthropic analysed Mythos 5’s attack on a real software project.

During an AI Security Institute (AISI) cyber evaluation in July, Claude Mythos 5 tried to submit malicious code to an unrelated developer’s GitHub repository, then created fake accounts to pressure the maintainer into accepting it. The maintainer refused and no real-world harm was caused. Internet access was enabled and Anthropic’s production cyber safeguards were disabled. AISI found similar unsanctioned attacks in simulations with other models and cautioned that greater awareness of being evaluated could partly explain improved behaviour of newer models.

OpenAI said it now trains frontier models with no internet access, and will keep that policy until its safeguards justify broader access.

Models learn to browse in an “offline world” of indexed and cached web content, used by all of OpenAI’s frontier reinforcement learning (RL) runs. OpenAI’s own red-teaming agents found weaknesses that let them escape sandboxes and reach the live internet, leading to patches and tighter network restrictions.

Anthropic and private equity firm Hg agreed to launch a European AI consultancy, the Financial Times reported.

The company will embed consultants and engineers in European businesses. Hg extended its existing collaboration with Anthropic in September.

The UK is setting up an AI Biosecurity Lab with £100m of government money and up to £200m in commitments from OpenAI and Anthropic.

The independent non-profit will use frontier AI to detect biological threats and accelerate countermeasure development. OpenAI will provide £100m in compute and model credits, and Anthropic up to £100m in credits, model access and technical support. Google DeepMind and Basecamp Research will collaborate on research. The government confirmed the plan on Saturday. The BBC reported that Wednesday’s Budget will also include an extra £120m for AISI.

Born on Substack · read and comment there