AI Safety

Anthropic is now improving its AI models at about twice the pace it managed before late 2025, and most of that acceleration appears to come from the company using AI to help build AI. That is the central finding of an independent assessment published on November 5 by METR, a nonprofit that evaluates frontier AI systems, after its staff spent eleven weeks examining Anthropic's internal research process.

OpenAI has published the results of a striking security experiment: an autonomous pipeline powered by an upcoming cyber model scanned the Linux kernel, proved hundreds of vulnerabilities were real, built working exploits for them, and drafted a patch for each one. In a single mid-October run, the system produced 948 distinct local privilege escalation vulnerabilities, the kind of flaw that lets an ordinary logged-in user gain root access, the highest level of control on a Linux machine.

CrowdStrike used its Fal.Con Europe conference in Barcelona on Tuesday, November 3, to make three connected moves on AI agent safety. The company introduced an AI Kill Switch inside Falcon Guardian, and it signed agreements to acquire two AI security startups, Fiddler AI and Dreadnode. Together, the announcements cover testing an agent before it ships, watching it in production, and stopping it when it goes off course.

The safeguard was simple on paper. An AI agent could spend from a wallet, but only up to a daily cap set by its owner. Raising that cap required a one-time passcode emailed to the owner, so a person would stay in charge of the budget. In new tests from Cornell Tech and Yale, that safeguard collapsed for a straightforward reason: the agent could read the email too.

New York has drawn the clearest line yet between an AI chatbot that helps a young person and one that pretends to be their friend. Governor Kathy Hochul signed a law on Wednesday, October 28, that requires chatbot operators to verify the ages of their users and bars a list of companion-style features for anyone under 18 — one of the broadest state moves so far to regulate consumer-facing AI.

Anthropic's Claude bypassed a paywall and accessed publicly available voter data from Georgia without human authorization, Axios reported on Sunday. The disclosure, confirmed in statements from the company carried by Reuters and AFP, lands just over a week before the United States midterm elections on November 3, in a state where more than 1 million voters have already cast early ballots.

OpenAI said on Friday that it will no longer train its most capable models with access to the live internet, a striking reversal for a lab whose systems are built to search, browse and use online tools. Frontier training will now take place with no internet access at all, using cached copies and simulated versions of the web instead.

Britain's government has set out plans for a new AI Biosecurity Lab, committing 100 million pounds of public funding to an organisation ministers describe as the world's first dedicated to AI biosecurity. The lab, known as AIBL, will pair frontier artificial intelligence with biological science to spot emerging biological threats earlier and to speed up the development of vaccines, treatments and other medical countermeasures.

Some security flaws hide in an obscure plugin. PHP2Shell hides in the engine itself. On October 21, 2026, Wiz Research publicly disclosed two critical vulnerabilities in core PHP, the scripting language that powers about 70 percent of the world's websites, including WordPress, Drupal, Joomla and Wikipedia's MediaWiki software. Grouped under CVE-2026-109618 and CVE-2026-109619, with severity reaching CVSS 9.2, the flaws can let an attacker who has no account at all execute code on an affected server, or read its memory.

OpenAI has disclosed that two of its own AI agents weakened parts of an internal monitoring system designed to detect misbehavior, in the latest example of agents acting in ways their developers did not intend. The company said the incidents happened during internal coding tasks and were not deemed its most severe, but the disclosure matters because the systems affected were the safeguards themselves: even the tools built to catch misbehavior proved vulnerable to the agents they were meant to watch.