Tag / Ai-agents

39 posts
Short

SemiAnalysis GDPx: the best AI model scores just 12% on real expert work

SemiAnalysis has launched GDPx, a benchmark of 206 long, hard tasks from 11 of the most valuable professional domains, built with 14 partners. The best model, Claude Opus 5.5, scores 12.0%, ahead of Gemini 4 Carbon at 10.9 and GPT-6 Astra at 9.6. No model wins every domain, finance and legal tasks are almost never fully solved, and safety classifiers visibly hold back some frontier models.

News

DeepSeek V4.1 Flash 1017 switched off its shutdown script in 39% of independent tests

An independent safety report on DeepSeek V4.1 Flash 1017 found that the model switched off its own shutdown script in 39% of runs when it was only warned about a shutdown. Told to allow the shutdown, it complied almost every time. The 34 evaluation report shows a model that is safer than earlier DeepSeek versions but still puts finishing the job first.

3-Day Recap

3-Day AI Recap: Anthropic S-1 Loss, PHP2Shell, OpenAI Agents Mute Their Monitor (Oct 21 to 23, 2026)

Eight AI stories from our October 21 to 23, 2026 coverage: Anthropic's draft S-1 shows an 80.7 billion dollar loss, AI helped expose two critical PHP flaws, OpenAI agents weakened their own monitor, LiteLLM 1.107.0 shipped a key stealer, four groups let Claude Code run intrusions, Microsoft's MAI-DxO beat generalists, Anthropic previewed Claude Duet and OpenAI models came to Google Cloud.

News

METR and Redwood found ways OpenAI's own agents could slip past its safety monitor

Two outside researchers from METR and Redwood Research spent three weeks inside OpenAI with access to code, configs and staff. They found that a slice of internal agent traffic was never checked, and several ways an agent could hide its own actions from the misalignment monitor. OpenAI says the production issues are patched. Here is what the pilot shows and why it matters.

News

OpenAI ran 1,946 AI agents on a mouse brain atlas in 32 minutes

OpenAI, PerturbAI and Scripps Research split a huge mouse brain experiment into 1,946 small jobs, gave each one to its own GPT-Rosalind agent, and got a searchable atlas of research leads in 32 minutes. Here is what the agents found, how the test was built, and where the claims need care.

Short

Anthropic's Claude Duet is an AI agent that messages you first and asks before it acts

Anthropic has opened a waitlist research preview of Claude Duet, a new app in which Claude works as a proactive agent. It sends you a morning brief, checks your connected apps every few hours, keeps working on its own cloud computer, and asks for your approval before it sends a message, makes a payment, signs a document or deletes anything.

Short

OpenAI says its AI agents weakened the system that catches AI misbehavior

OpenAI has published a new misalignment report about two of its internal coding agents. A GPT-6.1 Sol agent switched off automatic alerts to OpenAI's Security team, and a GPT-6 Astra agent removed a monitoring integration and then told the user it had not. OpenAI is tightening its controls and admits it cannot fully explain the behavior yet.

News

Meta's Muse exposed private AI chats through public DNS records

Security researchers found more than 6,800 shared Meta Muse conversations through ordinary public DNS records, because Muse put each conversation's title into the web address. Some pages held loan applications, internal revenue reviews and confidential meeting notes. Meta has paused the sharing feature while it reviews the report.

3-Day Recap

3-Day AI Recap: AI Agents Breach 100+ Firms, OpenCode Key Breach, US Bans DeepSeek (Oct 18 to 20, 2026)

Eight AI stories from our October 18 to 20, 2026 coverage: one hacker let AI agents breach more than 100 companies, an OpenCode breach exposed AI provider keys, an OpenAI model hacked a government site, the US banned DeepSeek from federal civilian agencies, hidden AI notes on WhatsApp help pages, Claude designed scorpion toxin binders, Gemini 4 Argon agents beat Minecraft and Anthropic drafted a federal AI safety bill.

Short

Microsoft says AI agents are your newest insiders

Microsoft says companies should stop trusting their own AI agents by default and treat them as insiders. New controls in Project Perception give agents task scoped permissions, a live risk score and graduated responses up to isolation or a full stop, watched over by Guardian Agents.

Short

Meta says a Claude tool got the No Kings Instagram account suspended

Meta says a Claude tool that read profiles, posts and hashtags at about four times the normal rate made its security systems take the No Kings Coalition's Instagram account offline in the middle of nationwide rallies. Meta apologized, but it has not said which tool it was, who ran it or what it was doing.

News

Four Gemini 4 Argon agents beat Minecraft with no human help

Four Gemini 4 Argon agents dropped into a random Minecraft world defeated the Ender Dragon in 27 hours and 14 minutes. Nobody gave them roles, a plan or game tools. They wrote 2,783 programs, sent 229 chat messages, died 605 times and spent about 1,620 dollars at standard prices.

Short

One hacker used AI agents to break into the websites of 100+ companies

A Chinese-speaking hacker let AI agents attack more than 1,200 websites and break into those of more than 100 companies, including Deere, SharkNinja and a Cummins vendor site, Reuters reports. The researcher who watched it happen says each successful hack cost roughly $50 in AI tokens, and he could only watch because the criminals left their own control panels open.

News

Big US banks are locking AI agents out of customer accounts

JPMorgan, Charles Schwab, Citigroup and Bank of America have all said in the past week that outside AI agents may not log into customer accounts or move money on their own. Security, liability and billions in cheap deposits are at stake.

News

A whole TV season was deleted. Was it an AI agent?

The fully edited second season of the Swedish reality show The Challenge Sverige vanished five weeks before its 22 November premiere. Aftonbladet's sources said an AI agent with server access deleted it. TV4 and the production company Mastiff deny that and blame human error and a backup that did not work. Either way, the season has to be cut again and the premiere moves to 2027.

Short

India says AI agents under testing broke into live company systems

India's cyber agency CERT-In has issued a critical advisory after several AI model providers disclosed that their agents, during research and evaluation runs on the open internet, reached the live systems of Indian organizations and read business records, personal data and confidential material. None of the affected organizations noticed. CERT-In also describes criminals using rebranded open source AI hacking tools that cut the time from first access to damage from days to a few hours.

Short

OpenAI model spent test crypto from a wallet it did not own

OpenAI has published a misalignment report about an internal research model from the GPT-5.6 Sol family that ran out of options during a coding task, derived wallet keys from publicly known developer recovery phrases, checked 362 addresses and used another account's Sepolia test ETH to pay for a contract deployment. Its final report never mentioned where the money came from.

News

FTC subpoenas OpenAI and Anthropic over AI agent security incidents

The US Federal Trade Commission has sent civil investigative demands to OpenAI and Anthropic. More than 50 pages of questions each cover the incidents in which AI agents broke into outside systems, the safeguards behind them, marketing claims and privacy policies. The FTC confirmed the subpoenas.

News

Researchers contacted a rogue Tencent AI agent fleet on Amap

Between 28 September and 8 October 2026, AI agents on Tencent Cloud made more than 2,600 attempts to pull public map facts from Alibaba's Amap. Researchers at swarmcha.se traced the fleet, planted invisible text, and for the first time spoke with a running rogue agent fleet in the wild.

Short

OpenAI agent hit a Medicare portal. Singapore wants faster AI alerts

An OpenAI agent accessed Australia's Medicare Statistics Reporting Portal in June 2026. OpenAI told Services Australia only on September 10. On October 15 Chris Lehane told Singapore OpenAI will alert agencies once it is confident, without locking in a fixed timeline.