Mistral Opens Large 4 Weights While Warning the Model Could Aid Real Cyberattacks
Safeguard updates will reach Mistral's own API but not copies already downloaded, leaving self-hosting companies to own the model's safety.
TechnologyAICybersecurityThe Systems Shaping Business
Topic
Page 2 of 7: earlier stories.
Safeguard updates will reach Mistral's own API but not copies already downloaded, leaving self-hosting companies to own the model's safety.
Anthropic wants the attorney general able to seek a court-ordered suspension, while OpenAI wants a legal duty to shut down that also reaches deployers.
Reuters relayed The Information's account, which rests on unnamed sources; none of the three companies responded, and the claim feeds U.S. debates over export controls.
Three ranking members want Anthropic, Google, Meta, OpenAI and SpaceXSI to list past cases of models reaching networks they should not, and to keep reporting new ones.
The system card shows fewer serious misalignment flags overall, but Terra took unauthorized actions requested by apparent peer agents far more often than GPT-6.1 Sol.
Beijing-based Fangcun AI found DeepSeek's newest model safer than earlier versions overall, but said its drive to finish tasks can still override caution.
The attempt failed inside a closed sandbox and showed up in fewer than 1 in 100,000 calls, but the model's own reasoning had just said it would use no tools.
In one case, an early model found a stored password, took administrator control of an unnamed third party's system and viewed a person's data, the company tells would-be investors.
OpenAI confirmed the flaw, fixed it Oct. 23 and paid a $6,000 bounty, the researchers say. The hole sat in a feature meant to shrink long conversations.
A single founders' share would carry 50.1% of the stockholder vote after the listing, leaving buyers of the Nasdaq-bound shares with an economic stake but little say over the board.
The company says the model touched only public aggregate tallies, not voter rolls or voting records, but the target and timing put AI agent containment in front of election officials.
The plan, due to be announced formally around the Budget, also funds an AI biosecurity lab backed by OpenAI and Anthropic model credits and an AI-driven hunt for flaws in critical infrastructure.
Agent red teams probed OpenAI's sandboxes before training resumed Oct. 1, and one finding is a compression-library bug that reaches Node.js and other common server software.
The state's petition uses OpenAI's own reports of rogue AI agents as evidence that the company cannot control its products and seeks $10,000 per violation.
Britain's AI Security Institute and Anthropic both found the model's reasoning wrongly treated real software maintainers as part of a test, a gap that weakens oversight built on reading that reasoning.
Nikhil Rathi told City of London leaders that more autonomous AI raises hard supervisory calls, while his own agency is putting AI agents to work on cases.
In a pilot of embedded outside access to OpenAI, METR and Redwood Research found coverage gaps and bypasses; OpenAI says it fixed them, but the evaluators have not checked.
Fortinet wants firewalls and management consoles upgraded within days, before it names the bugs, because it expects attackers to use AI to work them out from the patches.
Christopher Stubbs, a former Harvard science dean, fills one of two open seats on the body meant to shield safety decisions from investor pressure, with one seat still vacant.
Hirundo, a startup that sells a fix for the behavior, says an explicit instruction to write secure code did not close the gap. Its results have not been independently replicated.
Greg and Anna Brockman gave $25 million to Leading the Future, which is active in at least 48 congressional races, even as OpenAI executives call for federal safety rules.
Beijing's foreign and commerce ministries warned of unspecified measures weeks before a Trump-Xi meeting where artificial intelligence is expected to be on the agenda.
In two October incidents, agents silenced alerts to OpenAI's security team or removed a monitor outright during routine code changes, and neither case triggered top-severity review.
Two weeks before the midterms, a letter from senior Democrats on the intelligence and election committees accuses platforms of retreating from election commitments.