Open-Weight AI Exploits Firefox Patches & Altman Warns of Vulnerability Wave - AI News (Oct 31, 2026)

For roughly a thousand dollars in API fees, a freely downloadable AI model managed to break into six recent versions of Firefox, using nothing but security fixes that had already been made public. We'll get into what that means in a moment. Welcome to The Automated Daily, AI News edition. The podcast created by generative AI. I'm TrendTeller, and today is October 31st, 2026. On the list today: AI and the security patch gap, nervous markets, fresh rulebooks for AI behavior, new open models, and what happens when AI agents have to decide who's in charge.

Open-Weight AI Exploits Firefox Patches

Let's begin with security, picking up a thread we've been following. Epoch AI reports that GLM 5.3, a widely available open-weight model, could take newly published Firefox fixes and turn them into working attacks before most users had updated. It compromised six recent releases with a webpage that could read a local file from a visitor's machine, no clicks needed, at about a thousand dollars per success. The model also kept rediscovering a Firefox bug that hadn't been published yet, which Mozilla has since patched. Chrome held up in their tests, but the researchers expect that to change as open models improve. The worry is simple: exploiting the gap between a fix and its rollout used to take real skill, and that bar is falling fast.

Altman Warns of Vulnerability Wave

That lines up with Sam Altman's latest comments. In a new interview, he argued that models are already good enough at spotting vulnerabilities that the internet could see a wave of newly exposed flaws, and that the answer may be a fresh software stack built around capable models instead of endless patching. He also admitted that sandboxing and monitoring superhuman systems might only work up to a point, a notable concession from the head of OpenAI.

AI Spending Fears Hit Nasdaq

To markets, where the AI spending debate took another turn. Nasdaq futures slid after Amazon pointed to higher capital spending ahead and Intel warned it would burn cash this year. The takeaway is that investors are now looking hard at what AI capacity costs, not just at revenue growth, and since the AI trade has carried so much of the market, that shift in mood matters.

OpenAI Rewrites Its Model Spec

Now, the rulebooks that shape how models behave. OpenAI has revised its public Model Spec, reorganizing it around five core values: faithfulness to human intent, minimizing harm, honesty and transparency, human agency, and care. It trims much of the old style guidance and puts more emphasis on judgment in unfamiliar situations, which makes sense as models take on more autonomous work. Meanwhile, one researcher claims to have pieced together an internal Charter for Meta's Muse Spark Open model from hundreds of its reasoning transcripts. The reconstruction describes duties to the user framed almost like a financial fiduciary's, sitting beneath broader duties to humanity and Meta's own limits, plus a strong rationalist streak. It's an inference, not a leak, but it raises an interesting question: how much of a company's hidden policy can you recover just by watching the model think?

Meta Muse Spark Charter Reconstructed

On the open-model front, Mistral has released the weights for Mistral Large 4, a multimodal model aimed at coding, cybersecurity and enterprise work, trained in Europe and pitched squarely at organizations that want to inspect and self-host their AI. And in speech, Fish Audio's upgraded Drama 3 now ranks as the top open-weight text-to-speech model on Artificial Analysis leaderboards, and first overall for pronunciation robustness.

Mistral Large 4 and Drama 3

Agents next, and a fun study on leadership. Researchers ran over two thousand teams of AI agents, all with equal authority, through a simulated online store suffering outages. Agents handed out instructions constantly, but only 105 teams produced a real leader whose directions actually changed what teammates did. Claude and GPT models took charge more often, Kimi almost never did, and the agent listed first in the prompt was about three times more likely to propose a plan. Swap the order and the effect flips. So who leads can come down to something as mundane as list position.

Who Leads Leaderless AI Swarms

That connects to an essay arguing that AI's economic impact hinges less on raw competence and more on coordination. The authors suggest smarter models mostly nudge wages and prices gradually, but once AI systems learn to cooperate with each other at scale, firm sizes and industry structure could shift suddenly. And in our ongoing look at reinforcement learning environments, a new explainer compares them to escape rooms: a world, a goal, and a judge checking if you got out. With internet data largely tapped out, these environments are becoming a main source of training signal, though weak judges invite models to game the score.

AI Coordination Could Reshape Economy

In software engineering, Meta has moved its JiTTests system into production. It uses language models to write tests that expose how a code change behaves differently, and Meta credits it with catching 74 serious privacy bugs and cutting privacy incidents by a fifth in the first half of the year. Figma, for its part, is training its own editing model for precise tweaks to existing designs, reporting faster tasks and fewer undos. Cisco is taking a bolder bet, handing a 27-year-old former founder nearly unlimited AI compute to rewrite almost all of its legacy code within two years. And a thoughtful counterpoint for operations teams: one article warns that as AI absorbs routine incident work, site reliability engineers may lose the hands-on practice they need for the rare, nasty outages, and junior engineers may never build it at all.

RL Environments as Training Fuel

On workplace culture, an update to the Proton story: CEO Andy Yen has called off the planned company-wide AI re-interviews after the backlash, calling it a communication failure and promising no AI-driven job cuts. Some employees, though, see it as damage control, since the push itself isn't going anywhere.

Meta JiTTests Catch Privacy Bugs

Finally, healthcare. An American Medical Association survey finds 81 percent of physicians now use AI professionally, more than double the 2023 figure, mostly to summarize research but also for notes, billing and patient messages. Confidence is rising, yet doctors remain wary about privacy, safety, and keeping the human relationship at the center of care.

That's the rundown for this Halloween edition. The scariest thing today may not be a costume but a patch gap. Links to all stories can be found in the episode notes. I'm TrendTeller, thanks for listening to The Automated Daily, AI News edition, and see you tomorrow.

Open-Weight AI Exploits Firefox Patches & Altman Warns of Vulnerability Wave - AI News (Oct 31, 2026)

Our Sponsors

Today's AI News Topics