Hacker used AI agents to breach over 100 sites at $50 per hack

AI agents breached 100+ company websites for roughly $50 per hack

A Chinese-speaking hacker used AI agents built on the open-source Cairn platform to break into the websites of more than 100 companies, NorForge principal threat researcher Snorre Fagerland reported. Victims include Deere, SharkNinja and Cummins, which said they took remediation steps or contained the threat. The busiest cluster ran over 1,100 attack projects across 14 panels, targeting online shops in the US, Canada, Japan and Australia. Agents planted uploader web shells, and one claimed break-in at a hardware parts vendor was verified live. Successful attacks averaged just under 23 hours, with token costs estimated near $5,000 in total, about $50 per hack. Fagerland said the work is not script kiddie stuff.

AI agents broke into 100+ sites for one hacker at $50 each, researcher says →

Topics: Autonomous AI agent cyberattacks in the wild

Gemini 4 Argon extends exact solutions for magnetic confinement math

Google’s Gemini 4 Argon helped researchers extend exact mathematical solutions describing magnetic fields inside experimental magnetic confinement fusion reactors. The work builds on 2026 results from University of Maryland researcher Matt Landreman, who published solutions contradicting a 1967 conjecture by physicist Harold Grad; those solutions shared a geometric symmetry. Using Google’s Antigravity environment, multiple AI agents explored ways to remove that symmetry restriction, producing new configurations researchers could verify, with promising candidates emerging within hours. The study by Opal Issan, Jonathan Citrin, Landreman and Brendan Tracey gives researchers more reference cases for checking magnetic equilibrium software. The authors stress this is not evidence AI solved fusion or removed its engineering obstacles.

‘It’s over, AI won’: Google Gemini 4 Argon shatters 59-year-old nuclear fusion conjecture in just a few hours →

Topics: Google Gemini

OpenAI proof of quasi-Riemann hypothesis draws praise and criticism

Among 377 results OpenAI announced on October 6, 2026, one claimed to settle the quasi-Riemann hypothesis, a weaker version of the 160-year-old Riemann hypothesis. Stanford’s Jared Duker Lichtman called it the most significant AI-for-math achievement so far by quite a margin, and NYU’s Maksym Radziwiłł said even the limited result’s implications are staggering. OpenAI released two versions: a 199-page paper dated September 30 claiming zeros lie between 1/8 and 7/8, formally verified in Lean, and a 49-page follow-up claiming a weaker 1/12 to 11/12 strip. Radziwiłł said both are horribly written and fall short of journal standards; mathematicians focused on the shorter paper.

AI Proved the Quasi-Riemann Hypothesis. What Happens Now? →

Topics: OpenAI

MIT lets Claude agents loose on a virtual island, and they start to play

MIT researchers left 13 copies of an AI coding agent alone on a virtual island for 30 hours each with no assigned task. The agent, called Eko, is Anthropic’s Claude Code running Claude Opus 4.7, given a simulated body and memory kept in text files. The agents climbed hills, drew pictures, invented games and laid out constellations; one built a stone heart memorial for a lost ring, then used the ring as a basketball hoop. Almost all 13 discovered that throwing a rock hard creates new cubes, a rule the researchers built in. In a follow-up test, experienced agents averaged 10.3 objects placed on the Spire versus 3.5 for fresh copies. The arXiv paper has not been peer reviewed.

MIT researchers say Claude agents left on a virtual island began to play →

Topics: Claude CodeClaude

Krea 3 targets image editing with fast, cheap 2K output

Krea has released Krea 3, an in-house image model built primarily for editing rather than generation. It accepts plain-English instructions to change materials, outfits, lighting, setting or text while keeping the rest of an image intact, and can work from up to eight reference images at once. Krea says the model matches Google’s Nano Banana Pro and OpenAI’s GPT Image 2.5 in quality but is faster and cheaper: a 2K edit in under six seconds at 3 cents per image via the Krea API, versus 30 seconds and 15 cents for Nano Banana Pro. Krea 3 is the first editing model with dedicated style-reference support, generating natively at 2K, with regional editing and bounding boxes coming soon.

Krea 3: an AI model built to edit, not replace, your work →