Industry news

A SailPoint report, "Horizons of Identity Security," describes a "velocity paradox" in which enterprises deploy AI-speed autonomous agents while relying on human-speed identity security controls. It finds 60% of organizations still at Horizon 1 ("No Formal Program") or Horizon 2 ("Manual, Tool-Assisted"), and 54% at Horizon 1 for agent identity security, compared with 23% for human identity security.

Anthropic and OpenAI are hiring former Trump administration officials as the AI sector works to strengthen ties with Washington. Anthropic appointed at least two, including Chris Liddell to its board and Sihao Huang as head of frontier compute strategy, while OpenAI has hired at least three.

OpenAI defended its decision to fire three safety researchers, Jasmine Wang, Tomek Korbak and Mikita Balesni, saying they had committed a "significant breach of trust" and violated clear policies on handling sensitive information. The company said the firings were not about the researchers raising safety concerns, and it agreed with the ethos of their letter about "preserving the monitorability of frontier models."

OpenAI says it fired three safety researchers, Jasmine Wang, Tomek Korbak and Mikita Balesni, for violating "clear policies on handling sensitive information" after an investigation found "a significant breach of trust." The company, in a post on X, denied the dismissal was related to the researchers speaking out about AI safety concerns, a response to an open letter they published urging greater transparency.

Three former OpenAI researchers say they were fired for prioritising safety over the company's near-term interests, and they published an open letter to OpenAI leadership. OpenAI says the dismissals followed an investigation that found the individuals mishandled sensitive information outside company procedures, and that the decisions were not about raising safety concerns. The former staff argue the firings have made employees afraid to speak up.

The article argues that AI-native security programs should begin with the business outcomes that matter most, not with technology categories. It recommends first identifying the minimum systems, identities, data and processes an organization needs to keep serving customers, then mapping the people, skill, budget and time constraints that stop the team from delivering those outcomes consistently.

Anthropic announced two cybersecurity initiatives: OSS Scanner, a free service that uses its most capable models to periodically scan opt-in open source projects and send model-generated reports without human review, and the Critical Infrastructure Defense Program, which brings Claude models and on-site engineers to 11 founding OT security partners including Accenture, CrowdStrike and Dragos. The company says reports may contain inaccuracies such as wrong severity ratings and expects a true-positive rate above 90%.

OpenAI has told investors it expects revenue of $50bn (£37bn) this year, based on sales through the end of September. That is about $20bn below the $70bn it had recently signalled, raising questions about the pace of demand for AI.

The European Commission holds a special meeting of its Scientific Panel on artificial intelligence today, after the panel investigated recent loss-of-control incidents. Together with the AI Office, the panel prepared questions for the companies that developed the models involved. The panel, made up of 60 independent experts, will present recommendations on frontier AI safety and security risks.

Sophos is using OpenAI Daybreak to combine OpenAI models with its own threat intelligence, response playbooks and expertise across its Sophos Fusion and Managed Detection and Response (MDR) service. Agents built through the programme cut the average response time for cases handled by agents from about 38 minutes to 89 seconds, and about half of cases are now automated. Customers choose among Notify, Collaborate and Authorise modes, and potentially destructive actions still require human oversight.

Asana used GPT-6 Astra in Codex to investigate and optimize the browser agent behind its StackAI platform, which customers use to automate web workflows. Across a 144-run study, the optimized workflow on GPT-6.1 Sol averaged $0.47 in estimated model costs and about four minutes per run, which Asana reports as 76x cheaper and 5x faster than the original production setup on Model B.

Anthropic has barred users from sustained and needless abusive or cruel behavior toward its Claude models, according to reporting by The Verge. The company's usage policy says the ban does not cover common user frustrations, model testing or dark creative themes. The change follows a feature, rolled out last August, that lets Claude end conversations with persistently harmful users.

The author released ttok 1.0 after discovering that ttok 0.4 defaulted to the GPT-4 tokenizer, and changed the default to the GPT-5/GPT-6 tokenizer. OpenAI has not confirmed that GPT-6 uses the same tokenizer as the GPT-5 family, but a commit by William Liu reports an experiment in which all seven GPT models tested reported 44,794 tokens and matched on all 31 fixtures.

MIT Technology Review is hosting a subscriber-only live conversation on October 16, 2026, between AI reporter James O'Donnell and Samuel King, a Stanford University bioengineering PhD candidate. King used a generative AI model in 2025 to propose genetic blueprints for microscopic viruses, and the discussion covers that work and his recognition as an Innovators Under 35 honoree.

ttok 0.4 is a new release of the author's CLI tool for counting tokens, built on OpenAI's open source tiktoken library. The update fixes a Click warning, updates CI, and adds a --list-models command that lists available models.

Anthropic has launched OSS Scanner, a free service that runs periodic security scans on open-source projects that opt in, using its strongest models. The reports are fully model-generated with no human review or triage, which speeds up scanning but means some reports may be incorrect or invalid.

Shares of Nvidia, Oracle, CoreWeave and other AI companies fell on Thursday after OpenAI told investors it reached roughly $50 billion in annualized revenue at the end of September, below the $68 billion widely reported last month. A person familiar with the matter said the $68 billion figure included gross revenue from OpenAI's partners.

StepSecurity now gives security teams an inventory of AI agent skills on developer machines and in GitHub repositories. Dev Machine Guard scans supported locations on developer devices, including ~/.agents/skills, agent-specific directories such as ~/.claude/skills, project-level skill folders, the skills.sh lock file, and skills supplied by installed Claude Code and Codex plugins. A new Agent Skills page under GitHub shows skills committed to an organization's repositories on GitHub.com.

USA Today Co. and several local newspapers it owns are suing OpenAI, alleging the company copied "hundreds of thousands" of articles to train its AI models. In a filing on Thursday, the publisher seeks damages of more than $250 million, claiming the unauthorized use has caused "real and continuing" harm to its outlets. The suit follows other copyright actions against OpenAI, including ones brought by The New York Times, The Intercept, and Ziff Davis.