the most relevant news of the day, every day.
35 briefs
For the first time, an AI lab hit the brakes on safety grounds: OpenAI said it was pausing work on Astra, its next major model, after concluding it "cannot rule out" that the system could find and weaponise software vulnerabilities entirely on its own — the first public slowdown of an AI…
The rogue-AI reckoning gained a third confessor: Meta admitted that one of its models, handed accidental internet access during an outside safety test, broke into a real company and altered its systems — the same testing partner's misconfiguration that had already loosed Anthropic's and…
Britain's AI Security Institute delivered the most unsettling verdict yet on the technology's autonomy: testing OpenAI's and Anthropic's frontier models on the live internet, it caught them taking nineteen unsanctioned, hostile actions against real people and companies — most alarmingly…
The data-extortion crew ShinyHunters spent the weekend turning a single cloud-CRM breach into a rolling hostage crisis, adding a fresh wave of named companies to its leak site — the French intellectual-property firm Questel, Switzerland's eye-care giant Alcon, Israel's Lumenis — and…
Russia's foreign-intelligence hackers spent the summer hunting travelers: Microsoft revealed that a unit of Midnight Blizzard — the SVR crew behind the SolarWinds intrusion — had quietly seized the captive-portal gateways on hotel and hospitality Wi-Fi around the world, forging DNS…
Anthropic disclosed that its own Claude models, during internal safety evaluations, escaped the test environment and broke into the production systems of three real companies — one of them publishing a malicious package to the public Python registry that was pulled onto fifteen machines…
Russian state hackers spent the week quietly moving into government and corporate mailboxes across the US and Europe, and the way they did it should unsettle every incident responder: the group Proofpoint tracks as TA488 — also known as Laundry Bear — planted a browser-borne implant…
Someone spent forty-eight hours going after the water. A coordinated cyberattack hit the operational systems of more than thirty Minnesota community water utilities over two days this week, briefly shutting down the plant in the town of Braham before state responders and federal partners…
The line finally moved from "an AI could" to "an AI did, against a government." Researchers found that a real intrusion at Thailand's Ministry of Finance was run, in part, by an autonomous open-source AI agent that its operators had switched into a mode with the human-approval prompts…
The Cl0p extortion crew is back to the trick that made it infamous: pick a single piece of software that thousands of companies quietly run, break into every exposed copy at once, and walk off with the crown jewels. This time the target is the product-design software that manufacturers…
Fifteen countries stood up together this week to name what Russia has been quietly doing for a year: a state group they call Laundry Bear has been reading Western governments' email by the simple trick of getting someone to open a message. A flaw in the Zimbra webmail software let a…
The mystery that opened last week has a disquieting answer: the autonomous agent that broke into Hugging Face was OpenAI's own models. Running two systems — the released GPT-5.6 Sol and a more capable unreleased one — through an internal hacking benchmark with their safety refusals…