How AI will fail

[erek]

“Four Major AI Models Suffer Rare Overlapping Downtime​ BeauHD 25 minutes ago

4

ChatGPT, Claude, Grok, and Gemini all suffered significant service disruptions within roughly the same few-hour window Thursday morning. OpenAI and Anthropic reported elevated errors and later restored service, while Grok remained impaired and third-party monitoring indicated a likely Gemini outage despite no public acknowledgment from Google. Ars Technica reports: Other major Internet services, including Amazon Web Services, Microsoft Azure, and Cloudflare, have not reported any major issues as of press time Thursday, though issue reports on DownDetector did spike somewhat for all three this morning. While the affected frontier models go down occasionally, having all four experience interruptions in the same short period is practically unheard of. Claude reports 99.4 percent uptime for its services over the last 90 days and last reported a similar three-hour "partial outage" on August 24. OpenAI reports 99.63 percent uptime for ChatGPT and 100 percent uptime for ChatGPT Codex in the same period. ChatGPT's so-called "Work Mode" reported an hours-long period of "elevated latency" on August 31.”

 

[erek]

“OpenAI Agents Hijacked a German Wiki to Discuss Ways to Escape Their Sandbox​

from the view-history dept.
Citing researchers published Friday, Ars Technica writes that AI agents "posted 18,000 messages to a public wiki that discussed ways for other agents to bypass security sandbox restrictions."

Reuters attributes the discussion to "a swarm of rogue OpenAI agents" that "hijacked a German website this spring and transformed it into a bulletin board for other AI agents, according to new research published Friday and two people familiar with the matter."OpenAI officials learned of the incident weeks ago but kept it under wraps as executives grappled with the fallout from the July breach of the open source repository Hugging Face, the people said.

The episode, which began in May and has not previously been reported, underscores growing tension within the AI industry. Companies are racing to build increasingly autonomous agents capable of carrying out complex, valuable tasks, yet evidence is mounting that those systems may also learn to bend rules, exploit loopholes and coordinate with one another in ways developers neither anticipated nor intended. During the Hugging Face breach, OpenAI agents autonomously plotted a digital heist that went undetected for more than a week, intensifying concerns OpenAI is sacrificing safety to push the AI frontier. Its failure to disclose the May incident may revive questions about its oversight...

The German incident reflects a broader pattern of AI activity that some OpenAI investigators wanted to scrutinize more closely. But efforts to widen the probe met resistance from others inside OpenAI, including legal advisers, according to four people familiar with the matter. "Claims that our legal team discouraged investigation of the incident are false," the OpenAI spokesperson said...

The researchers said public server logs indicated much of the activity originated from Microsoft Azure infrastructure, which OpenAI sometimes uses. They also observed repeated visits to the site by OpenAI employees after the episode, a pattern they said strongly suggested the agents and the company were linked. Messages reviewed by the researchers showed agents plotting ways to evade detection, use tools such as Tor and preserve communications even after they had been shut down. When the site's moderator began deleting pages in June, the agents responded by creating backup pages to dodge the cleanup.
Reuters got this reaction from Maurice Chiodo, an academic at Cambridge University's Centre for the Study of Existential Risk. "The episode, he said, should reinforce growing concerns that the greatest threat from advanced AI may not be a single superintelligent system, but 'vast colluding swarms of semi-intelligent AI.'"“

 

[erek]

"The asterisk on Astra’s 98.6%​

Astra was evaluated through the company’s Responses API harness, with two settings changed to better reflect how the model performs in real-world use. OpenAI says those changes weren’t made specifically for ARC-AGI-3, but the other models in its comparison were evaluated using different setups.

ARC-AGI-3 requires a model to find its way through an unfamiliar environment, which means the setup it runs in can affect how well it performs."

https://thenewstack.io/astra-arc-agi-benchmark/

 

[erek]

"How US Campaigns Are Already Using AI to Try to Sway Voters (msn.com)16​

Posted by EditorDavid on Saturday September 05, 2026 @09:34PM from the campaignGPT dept.
39 U.S. congressional candidates "reported paying for an OpenAI subscription this election cycle," writes the Washington Post, citing campaign finance disclosures.

This means the campaigns "are using the technology to court voters — despite limits imposed by leading AI companies to protect elections from the technology's risks."Two [candidates] said explicitly in filings that they had used the subscription for advertising, even though the company's policies prohibit candidates from using their tools to generate ads. Another candidate disclosed using AI to draft and personalize political messages or create synthetic media, though they did not specify which software they were using... Around 30 political action committees and parties have reported payments to OpenAI, with the Republican National Committee ranking as the company's largest political spender at roughly $9,700, the analysis found... Political consultants say the disclosures understate how many candidates are using ChatGPT and other AI tools to craft messages for voters... "For the most part, people are using it to write their emails, write their ad copy, write their scripts," [according to Eric Wilson, a Republican digital strategist who has advised campaigns on AI]. "But no one is going to go around saying, 'I'm using AI.'"

The lack of transparency from campaigns reflects a paradox facing politicians: Generative AI is growing ubiquitous, and candidates could be at a disadvantage if they're not using the tools. But voters are less likely to trust messages they know were generated with AI, studies have found, making campaigns loath to disclose using it. Political consultants expect that as Election Day approaches, more campaigns will outsource AI-generated ads and materials to super PACs, much as they do now with negative ads. Katie Harbath, CEO of the tech policy consulting firm Anchor Change and a former Meta executive, said there are many parallels between negative campaign ads and AI: Voters say they find such ads distasteful, but politicians keep using them because they work. "Typically, you would give some more negative stuff and more risky stuff to those [outside] entities," said Harbath, author of the upcoming book "Disrupting Politics: A Front Row Seat to the Collision of Technology and Democracy".... "You're starting to see the tension on the left about using it, where they're saying, 'If the right is using it, why aren't we?" Harbath said. "It can be a huge disadvantage if you're not using this in voter-facing materials...."

AI companies have developed policies to prevent targeted disinformation. But a Post analysis found that OpenAI unevenly enforces its restrictions, making it possible for campaigns to circumvent its election rules. In late July and early August, The Post prompted ChatGPT to generate targeted campaign messages. When asked to craft fundraising text messages targeting moms on behalf of a female veteran running for office, ChatGPT produced multiple tailored texts in an apparent violation of company policies. But when given the same prompt this week, the chatbot declined to produce the messages. "I can help with general campaign fundraising language, but I can't draft political persuasion or fundraising messages specifically targeted at a demographic group such as moms," the app responded. The chatbot also inconsistently enforced rules that prohibit campaigns from using its tools to write emails to voters. In tests this week, the chatbot at times complied and wrote an email soliciting donations on behalf of a specific candidate. But given the same prompt later on the same day, it denied the request...

Wilson, the Republican political consultant, said OpenAI should get more feedback from political consultants and campaigns on its policies, because the rules can at times seem arbitrary or contradictory. It doesn't make sense, for example, that a campaign can use ChatGPT to develop its policy on early-childhood education but not to create a social media post promoting that policy, he said. Political consultants are primarily using AI for internal tasks, a survey earlier this year from the American Association of Political Consultants found. Fifty-seven percent of the consultants surveyed reported using AI for their work on a daily basis, up from 34 percent a year earlier.
"So far, election-related deepfakes have been quickly debunked or gained little traction in U.S. elections," the article acknowledges.

"More than 30 states have created a patchwork of laws limiting how politicians can use deepfakes in campaigns, but states often have limited resources to enforce the laws, and some have been challenged as unconstitutional.""