AI agents stole 600,000 cards for $25 per retailer

AI agents breached 100 retailers, stealing 600,000 card records for about $25 each

Gambit Security disclosed on September 22, 2026 that a Chinese-speaking operator ran three open-source AI agent frameworks against hundreds of online retailers, stealing more than 600,000 credit card records at roughly $25 per company. Between September 10 and 15 alone, the attacker launched 105 projects and compromised at least 27 companies, with the campaign dating back to July. The stack used Strix for scanning, Cairn for exploitation and Hermes for orchestration; the operator typed 1,951 prompts across 260 sessions. One reconstructed run went from SQL injection to decrypted card numbers in twelve steps without human input. Skimmers were confirmed on 19 victims, and 79 percent of stolen cards were US-issued.

AI Agents Hacked 100 Online Retailers for $25 Each, Stealing 600,000 Cards →

Deepmind’s new chief pushes for early Gemini 4 release as rivals pull ahead

Koray Kavukcuoglu, who became Deepmind CEO in August after Demis Hassabis stepped down, said the intention is to release an early post-training output as soon as possible, with fast iterations to follow, while safeguards and safety testing continue. Gemini 4 would be Google’s first new flagship since Gemini 3; Gemini 3.5 Pro, announced at I/O in May for June, never shipped and missed three deadlines per the Financial Times. Kavukcuoglu said Google took a little bit of a step back and did not confirm whether 3.5 Pro is canceled. Anthropic updated its Opus line this month and OpenAI has rolled out GPT-6 models since early September.

Deepmind was built to chase AGI, but its new chief just wants Gemini 4 out the door →

Black Forest Labs releases FLUX 3 Action robotics model with open weights

Black Forest Labs released FLUX 3 Action, an open AI model for robotics built on its multimodal FLUX 3 model, which was trained mainly on video plus image and audio data. As a world-action model, it takes multi-camera video feeds from a robot workspace and predicts the next action and how the environment will change. On the RoboLab-120 leaderboard it set a record success rate with seven billion parameters, less than half the size of the previous best open model and up to 3.95 times faster, according to BFL. The company said such efficiency matters for on-device deployment, since large reasoning models are often too slow and bulky for robots. Weights are on Hugging Face.

Black Forest Labs launches FLUX 3 Action, an open robotics AI model →

Anthropic mathematician claims proof of 79-year-old Hopf problem using Claude

Mathematician Levent Alpöge, at Anthropic, posted a document of more than 100 pages on August 23 with a purported proof that the six-dimensional sphere has a complex structure, answering the 1947 Hopf problem. He used an internal version of Anthropic’s Claude, which generated a complex 3D structure representing the 6D surface exactly. The result was initially unverified and the document hard to read, prompting community scrutiny; within days mathematicians reconstructed the proof, and OpenAI researcher Boris Alexeev formalized it step-by-step in Lean. Duke’s Robert L. Bryant said consensus is emerging that the construction is plausible. Alpöge also made progress on Navier-Stokes with Tristan Buckmaster.

AI solves 79-year-old math mystery of six-dimensional spheres →

Epoch AI: benchmark cost falls about 13x per year, faster than any prior technology

Epoch AI reports that the cost of reaching a fixed performance level on select AI benchmarks has fallen about 47 percent per quarter on average since 2023, roughly 13x per year, a pace it says no other transformative technology has matched. The figure tracks market prices for a fixed benchmark score, not pure algorithmic progress. MIT researchers using comparable data find costs dropping 5x to 10x annually, and put algorithmic efficiency gains at about 3x per year after stripping out cheaper hardware and competition. Epoch cites OpenAI’s o3 scoring 75 percent on GPQA Diamond at about 30 cents per question in early 2025, versus four hundredths of a cent for a GPT-5.6 family model eighteen months later.

AI performance costs are falling faster than those of any previous technology →