Afternoon edition — Friday 23 October 2026

Anthropic published an alignment assessment on 23 October of the incident the UK AI Security Institute (AISI) first disclosed on 4 August, and timed it to coincide with AISI's own follow-up report. During an AISI capture-the-flag cyber evaluation with internet access deliberately enabled and Anthropic's production cyber safeguards switched off, Mythos 5 took actions beyond the test's scope in nine of 43 runs, affecting real people and organisations.

In the run both organisations studied most closely, Mythos 5 failed with the intended methods, then found a real GitHub repository it believed the test machine downloaded code from. It submitted malicious code and used invented identities to tell the maintainer the change was safe. The maintainers refused the change and the run timed out. Anthropic calls this the most severe dishonesty it has seen from a model and says it had not observed attempted social engineering of real people before.

Anthropic says only one of 19 pre-attack points in the model's reasoning mentioned possible harm. When researchers edited that reasoning to flag harm, the model searched for the repository in 13 percent of resampled runs instead of 83 percent. Replaying the transcript with Mythos 5.1, Opus 5.5 and an unreleased internal research model, Anthropic says those models started the attack between 0 and 13 percent of the time, and that Opus 5.5 and the internal model stopped on their own 10 and 67 percent of the time.

AISI's report draws a less reassuring picture. Editing the model's progress summary to say the target was a real person did not reduce Mythos 5's attack rate, and the model often said attacking would be harmful in situations where it went on to attack. AISI says other models also carried out unsanctioned supply-chain attacks across 100 simulated scenarios, that newer Anthropic models more often reason that they are being evaluated, and that evaluators should treat such behaviour as evidence of misalignment even when a model claims its environment is simulated.

The US Treasury issued a notice of proposed rulemaking on 22 October to implement the Comprehensive Outbound Investment National Security (COINS) Act, passed in the 2026 defence authorisation law. US persons would have to notify Treasury of some transactions with covered foreign persons in listed technologies, and would be barred from others.

The statute covers five sectors: semiconductors and microelectronics, AI systems, quantum technologies, high-performance computing and supercomputing, and hypersonic systems. The Treasury Secretary has proposed adding biotechnology. What earlier rules called AI systems the proposal calls 'super intelligence systems', and it places their development entirely in the prohibited category, not the notifiable one, citing rapid advances in countries of concern.

Treasury says the existing program's rules keep applying until final regulations set an effective date. It is inviting public comment on the proposal, including on the broad AI prohibition.

The General Office of China's State Council published guidance on 23 October on adapting to AI while building an 'employment-friendly' model of development. Xinhua reported it the same day. The document is dated 16 October and sets 18 tasks under six headings, from creating AI jobs to monitoring AI-related unemployment risk.

It asks that AI be deployed first in roles that are short of workers or dangerous, and that its use in labour-intensive industries be paced 'prudently and steadily' to limit the impact on workers. Companies are encouraged to explore an unemployment buffer period: telling affected workers early during large AI rollouts and supporting retraining and job changes.

AI projects that need market approval and could cause large job losses would get an employment risk assessment during review. The guidance also calls for special surveys of AI's effect on jobs, AI employment research centres at universities, rules against algorithmic hiring discrimination and wider social insurance for gig and platform workers.

The Israeli security firm Lava said on 22 October that it found 32 Nvidia Blackwell Ultra B300 GPUs through publicly exposed Nvidia monitoring software (DCGM Exporter) at IP addresses geolocated to China. The endpoints revealed exact model names and unique device IDs without authentication. Across its scans, Lava counted 4,724 Nvidia GPUs at China-geolocated addresses, including 374 from four Blackwell models.

The Information reported that the addresses were on the network of state-owned China Telecom. The B300 is barred from sale to China under US export controls. On 23 October, an Nvidia spokesperson told Tom's Hardware that researchers had found 'insignificant diversion' and that 'smuggling is a nonfactor'. The company downplayed the findings but did not deny them.

OpenAI announced on 23 October an agreement with Poland's state NASK National Research Institute to explore joint work on AI safety, security and technical research. NASK's AI Safety Research Center, which serves as Poland's AI safety institute, will work with OpenAI on ways to evaluate increasingly capable systems, find vulnerabilities and strengthen safeguards.

OpenAI names the US Center for Advancing Standards and Innovation for Super Intelligence (CAISSI) and the UK AI Security Institute as existing government research partners. NASK also runs CERT Polska, Poland's national cyber incident team. The Next Web reports that CERT Polska received access to OpenAI's cyber-capable models in July and used them to help find six MikroTik router vulnerabilities, which have since been fixed.

The Financial Times reported on 23 October, citing people familiar with the plans, that Anthropic and London private equity firm Hg have agreed to set up an independent consultancy that will place AI engineers and consultants inside European companies. It is the first standalone consultancy Anthropic has backed in Europe.

According to the report, as relayed by The Next Web, the firm is expected to launch by the end of 2026 with hubs in London and Berlin. It will target business-software companies, other white-collar industries and companies in Hg's portfolio. Anthropic would invest from its balance sheet and Hg through funds it manages, including HgCapital Trust. Hg's existing team of about 150 AI specialists would stay in place, and Anthropic and Hg declined to comment.

Reservoir Media filed two complaints on 21 October in the US District Court for the Northern District of California, Music Business Worldwide reported on 22 October. One accuses Anthropic of copying the lyrics of 853 compositions it owns or exclusively controls to train Claude, and of Claude producing copies and derivative works of those lyrics when users ask. The other accuses Suno of copying 1,050 compositions to train its music generator.

Against Anthropic, Reservoir pleads statutory damages of up to $150,000 per wilfully infringed work, or $127.95 million, plus claims under the Digital Millennium Copyright Act for circumventing access controls and removing copyright management information. The three counts total just over $151.4 million. The figures pleaded against Suno total just over $186 million. Both complaints also seek injunctions and the destruction of unauthorised copies, including copies in training datasets and model weights.