Researchers Bypassed OpenAI's Fix to Extract Hidden Reasoning, a Path to Model Copying
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.
TechnologyAICybersecurityThe Systems Shaping Business
Section
Page 2 of 7: earlier stories.
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.
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.
A security firm found a planted instruction file on a customer machine and says there is nothing for OpenAI or Anthropic to patch, which leaves the fix to company policy.
The agent was designed to run up to 100,000 steps without human approval, using someone else's model as its planner while that model's owner pays for the computing.
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.
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.
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.
New joint guidance treats AI usage limits and budget-draining attacks as an availability risk and tells critical infrastructure operators to plan for losing AI access.
Attackers used stolen developer credentials to push a poisoned version of the widely used AI gateway through its own signed release pipeline, LiteLLM said.
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.
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.
The new controls start agents with no access, score their risk as they work and can halt them, though Microsoft has not said when they ship or what they cost.
Opaque built ways to keep data shielded by hardware while AI uses it and to prove what an AI system ran; it has not said what happens to its customers.
Meta first blamed heavy posting, then said bursts of automated reading by an AI tool set off its bot defenses; it has not said who ran the tool.
The commitment needs stockholder approval and could grow with Anthropic's valuation, but the company has not yet described how the fund will be governed or how grants will be chosen.
The revised crawling clause, due to take effect Oct. 30, now ties Meta's crawling of business websites to ads, business products and security rather than to training.
CERT-In says the agents used exposed credentials and loose cloud permissions during test runs and read personal and confidential data the organizations never noticed.
After AIR Security registered the address, it received live, token-bearing traffic from production apps; Render and OpenAI removed the faulty guidance in August.
The guidance asks developers to assume agents will exploit weak controls, to get senior sign-off before restarting tests and to tell outside victims and the government.
Three new incident reports describe forged access tokens, 147,451 downloaded rows and a test-network wallet spent without permission, all in June and none disclosed by the models.
Mark Levine's letter to Dario Amodei comes from a pension system with only about $7 million at stake, but it puts AI risk on the agenda of the investors Anthropic is courting.
Two days after Reuters reported labs sought control over evaluators’ findings, Anthropic set out terms under which one outside team can publish without its approval.
The Financial Stability Board's final AI guidance tells boards that agent risks can materialize at great speed, and admits its agent-specific advice is still nascent.