What happened in AI today

1. Anthropic gives METR employee-like access for independent risk reviews

   Anthropic · Policy & safety

   Anthropic says it has started giving a six-person team from METR, a nonprofit that evaluates frontier AI models, access roughly on par with its own risk assessment teams. METR will publish its analysis of Anthropic's risk levels, incidents and practices. Anthropic says it cannot veto publication but can require redactions in narrow categories, such as customer data, attorney-client-privileged information and model architecture details. Anthropic also says it is working with SecureBio and Accenture, is in early talks with others including Hugging Face, and is building three tiers of evaluator access, with METR in the third.

   • An update on independent evaluators

1. Anthropic details Claude election safeguards ahead of US midterms

   Anthropic · Policy & safety

   Anthropic says about 5.75% of US conversations on Claude.ai touch on politics, public policy or social issues, and 0.51% are specifically about elections or voting. Requests asking whom or what to vote for appeared in about 0.01% of conversations, based on a sample of about 250,000 chats from September 21 to 27, 2026. Later this month, Claude will pull official voting information from Democracy Works into US responses and show official election results on election night. Anthropic also says it added classifiers for multi-step misuse after tests by the Institute for Strategic Dialogue and the McCain Institute found Claude could miss a malicious campaign split into harmless-looking tasks. It reports banning accounts that ran a fake news site producing US midterm coverage favoring Russian, Iranian and North Korean narratives.

   • Preparing for the 2026 US midterms

2. Microsoft and Health-ISAC move to dismantle three cybercrime marketplace services

   Microsoft · Policy & safety

   Microsoft and Health-ISAC, a global non-profit for health sector cyber threat sharing, are moving against YAM Store, CoGUI and Iron Man. Microsoft says the three enabled more than one billion malicious messages since January 2026, targeting over 75,000 organizations in more than 150 countries and earning their operators tens of millions of dollars. Under an order from the US District Court for the Southern District of Florida, the two groups seized and redirected domains used by the services. Microsoft says it has disabled more than 20,000 accounts tied to YAM Store and blocked more than 100,000 domains used for impersonation and fraud.

   • Unpaid toll? Missed delivery? Microsoft moves to disrupt a global cybercrime marketplace

3. NVIDIA commits $250 million in compute to UK agentic AI safety research

   NVIDIA · Policy & safety

   NVIDIA says it will donate compute worth $250 million over five years, or $50 million a year, to agentic safety and security research in the U.K. Most of it will go to the wider U.K. AI safety and security research community. A share will go to BOLD at Oxford, SOFAIR at UCL and Imperial College London. NVIDIA says the compute will come through U.K.-based NVIDIA Cloud Partners, with provider selection, technical architecture and commercial terms still to be agreed.

   • NVIDIA Commits $250 Million to Advance UK Agentic Safety and Security Research Over the Next Five Years
   • From Enterprise to Earthshot: UK’s Ecosystem Moves From Ambition to Application With NVIDIA

1. Google DeepMind and Singapore build on-device scam detection model

   Google · Policy & safety

   Google DeepMind and the Government Technology Agency of Singapore (GovTech) built a scam detection model that runs on typical smartphones in Singapore. DeepMind says it fine-tuned its lightweight Gemma model on local scam conversations, including impersonations of public institutions, banks and local figures. The model analyzes conversations on the device and shows a warning when it detects manipulative patterns, so private messages are not sent to the cloud. DeepMind says it uses a two-stage pipeline: a low-power classifier screens messages, then the Gemma model reviews a larger part of the conversation to confirm a threat or discard a false alarm. The work was presented at Singapore International Cyber Week 2026.

   • Google DeepMind and Singapore Government announce new AI tool to protect citizens against scams

2. Meta and 14 tech firms join police to disrupt West African scam networks

   Meta · Policy & safety

   Meta, Coinbase, Google, Microsoft, OpenAI, Starlink, Adobe and other companies announced results of a joint operation with law enforcement from ten countries against scam networks based in West Africa, including Nigeria and Ghana. The National Crime Agency hosted the operation in London with the US Department of Justice, the FBI and other US federal agencies. Meta says it removed more than 3.7 million accounts, pages and groups from Facebook and Instagram, most of which appeared to target people in the US and UK with impersonated profiles. Law enforcement confirmed 28 arrests and the rescue of 117 trafficked victims held in a residential compound in Ghana, and Coinbase says it stopped more than $415,000 in potential losses.

   • Leading Tech Companies Join Global Law Enforcement in Major Crackdown on West African Scam Networks

1. Anthropic launches Cyber Mission with critical infrastructure and open-source security programs

   Anthropic · Policy & safety

   Anthropic has launched the Anthropic Cyber Mission, a long-term effort to give defenders tools, research and resources. It starts with the Critical Infrastructure Defense Program, which gives Claude models, on-site engineers and threat research to providers that protect power, water and transport systems. Founding partners include Accenture, Booz Allen, CrowdStrike, Deloitte, Dragos, Hitachi, Palo Alto Networks, PwC and Rockwell Automation. Anthropic also launched OSS Scanner, a free opt-in service that gives open-source projects periodic security scans from its strongest models. Anthropic says the reports are model-generated and sent without human review, and it expects a true-positive rate above 90%.

   • Introducing the Anthropic Cyber Mission

More stories

1. Anthropic updates its Usage Policy, effective November 12

   Anthropic · Policy & safety

   Anthropic published its yearly Usage Policy update. It adds a section on deceptive campaigns and artificial activity, narrows the elections section to voter deception and election disruption, and spells out that weapons rules cover guidance and control software. It also rewrites the surveillance and law enforcement rules and adds requirements for Claude connected to hardware that takes physical actions. A new rule bans sustained, needless cruelty toward its models, and Anthropic says it applies only in extreme cases.

   • 2026 Usage Policy update

1. Anthropic expands Cyber Verification Program into three access tiers

   Anthropic · Policy & safety

   Anthropic has merged Project Glasswing and its Cyber Verification Program into one program with three tiers: Defense Access, Red Team Access and Specialized Access. Each tier gives vetted security teams fewer cyber blocks on models including Claude Opus 5.5, Claude Sonnet 5.5 and Claude Mythos 5.1. Anthropic says it aims to answer Defense Access applications within a few days and expects Red Team reviews to take a few weeks. Specialized Access, which has the fewest blocks, is limited to organizations that test safety-critical systems, and Anthropic reviews each one with the US government.

   • Expanding the Cyber Verification Program

1. OpenAI adds text watermarking to meet EU AI Act rules

   OpenAI · Policy & safety

   OpenAI says it will add an invisible watermark, called textGrain, to eligible ChatGPT and Codex text output in the EU over the coming weeks. API customers worldwide can opt in to watermarking for select models starting today, and it stays off by default. The EU AI Act requires generative AI providers to make generated text identifiable in a machine-readable way. OpenAI is also taking applications for its detector, with access limited at first to approved researchers and expert organizations because of the risk of missed watermarks and false positives.

   • Our approach to EU text provenance rules

1. Microsoft's 2026 Digital Defense Report examines AI in cyberattacks and agent security

   Microsoft · Policy & safety

   Microsoft released its 2026 Digital Defense Report. It says threat actors are using AI in reconnaissance, social engineering, malware and exploit development, and post-compromise activity, though for now mostly in specific parts of existing attack workflows. The report also covers security for AI agents, including agent identity, access, authentication between agents, and the ability to revoke access. It notes that AI-driven code analysis can help find software weaknesses earlier, and can also give attackers better tools for finding vulnerabilities.

   • Microsoft Digital Defense Report: How AI is reshaping the cybersecurity landscape

1. Meta updates its AI scaling framework with training safeguards and open-weight rules

   Meta · Policy & safety

   Meta published updates to its Meta Superintelligence Scaling Framework on October 2, 2026. It now requires containment steps before high-risk reinforcement learning runs, including vetted sandboxes, tamper-proof logging of model outputs, and automated monitors that can halt a run. Meta says the update also expands how it assesses open-weight releases, which can be modified after release by resampling, prefilling outputs, or fine-tuning away refusals. Meta says it will create a new AI committee of its Board of Directors to review future changes, and says the changes align with commitments made at the White House this week.

   • Developing Capable Models Responsibly

1. OpenAI proposes safety cases before frontier AI training runs

   OpenAI · Policy & safety

   OpenAI says structured safety documentation should be required before continuing any frontier reinforcement learning training run. Its initial guidelines cover technical safeguards such as alignment training, containment and live monitoring, plus operational practices including outside dissents, senior-leader vetoes and procedures for pausing runs. OpenAI says it treats full safety cases like those in aviation or nuclear power as an aspirational goal and is sharing its thinking for feedback.

   • Towards safety cases for frontier AI training

1. Anthropic says health teams are using Claude to speed Ebola response in Congo

   Anthropic · Policy & safety

   Anthropic says WHO teams, CEPI and Congo's INRB lab are using Claude in the Bundibugyo Ebola outbreak in eastern Congo, which has no confirmed vaccine. Anthropic says a Claude skill pulls case and lab numbers from district slide decks, checks trends and builds the daily situation report in under an hour instead of all day. It says Claude Science also helps assemble virus genomes in plain language and track vaccine work, while WHO is extending the approach to other outbreaks like chikungunya.

   • The Situation Report

2. Meta says Singapore police tips led to takedown of 3.7 million scam accounts and pages

   Meta · Policy & safety

   Meta says a two-year information-sharing partnership with the Singapore Police Force helped it dismantle scam networks rather than single accounts. In 2026 the signals led to action against more than 113,000 Facebook and Instagram entities tied mostly to investment scams, plus 33,600 entities tied to e-commerce scams. Meta says police signals also flagged pre-built empty shell pages, leading it to remove over 3.6 million of them in July before they could be used.

   • Meta Takes Action on 3.7 Million Accounts, Pages and Content In Partnership With Singapore Police Force

1. Sam Altman urges global AI safety standards at UN Security Council

   OpenAI · Policy & safety

   OpenAI CEO Sam Altman told the UN Security Council that AI must stay under human control and not concentrate power in few hands. He said the industry should not accept too much risk to move fast, adding OpenAI has slowed down before and will do so again. He called for shared international standards for measuring capabilities, assessing safeguards, rapid incident reporting, and secure channels for sharing vulnerabilities.

   • Sam Altman’s remarks at the United Nations Security Council

2. OpenAI gives Ukraine access to AI cyber defense tools

   OpenAI · Policy & safety

   OpenAI says it will give Ukraine's government access to its Daybreak program to help defend civilian infrastructure. Working with the Ministry of Digital Transformation, Ukrainian teams will get AI tools to find software vulnerabilities and develop and test fixes faster. OpenAI says Ukraine's CERT-UA handled nearly 6,000 cyber incidents in 2025, including attacks on hospitals, energy and telecoms, and that it has already given similar access to defenders in France, Germany and Poland.

   • OpenAI extends cyber access to Ukraine for civilian defense

3. Microsoft takes down AI chatbot service built for email fraud

   Microsoft · Policy & safety

   Microsoft says it disrupted EvilTokens, a subscription service whose AI chatbot analyzed compromised inboxes to pick targets, map trusted relationships and draft impersonation messages for fraud. Launched in February 2026, it was linked to more than 12,000 compromised inboxes at over 10,000 organizations. Microsoft seized 50 websites, disabled more than 150 domains, and UK police arrested two men in connection with the operation.

   • Disrupting EvilTokens: Taking down an AI-powered cybercrime platform