AI News Report, October 19: CAIRN AI AGENTS BROKE INTO 100+ COMPANY SITES AT ABOUT $50 A HACK

Security firm NorForge found more than 20 Cairn control panels left open on the internet, showing criminals running over 1,100 automated attack projects against online shops through the open-source AI agent platform, with agents working about 23 hours per break-in at roughly $50 a hack and almost no human help, Reuters reported. The agents broke into more than 100 company sites, and the attack logs name DeepSeek's V4.1 Flash family. The finding matters because it documents real-world criminal misuse of autonomous agents against businesses at near-zero cost.

It meters agent use per client with hard spending caps and keeps each client's memory separate, with early access this week. Pax8 says 47,000 IT providers use its marketplace, so watch your partner portal. A new Dynamic Agent Risk score can block writes, require human approval or halt an agent, and discovery now covers 60 local agents, Claude Code included. List the agents on your endpoints first. The 1017 update lifts Terminal-Bench 4.0 from 31.2 to 42.3 on DeepSeek's own tests and keeps a 1M token context. Its V4.1 Flash family is the one named in the Cairn attack logs. MiniMax says it runs at over 200 tokens per second and cuts agent task costs 70% from its last model. Weights come later, so test it by API for cheap agent loops. It rewrites prompts, tools and memory around a fixed model and gained up to 11.5 points on WorkBuddy-Bench, under Apache 2.0. The repo warns to restrict its network access before you run it. The paper reports 10.4% better results with 35.5% to 54.7% fewer tokens across 11 benchmarks. If token bills drive your agent costs, this is the research to watch. He argues one proxy you control gives you per-key logs, virtual keys and fast model switching, and he traces a cache miss in about 20 seconds. After the OpenCode breach, virtual keys look smart. KillSwitch hides lies in its spec and gives "ignore previous instructions" a real meaning, and the best model scored 66.8 on simple tasks. A fun look at where coding agents still break. Linear's Saarinen worries that products built by many people, each running agents, lose a single point of view. A good gut check if your team ships agent-written features fast. He says LLMs now produce real proofs, some in Lean running to hundreds of thousands of lines, and asks for public control of frontier tools. A sharp take on who owns AI-made knowledge. Not content with challenging the major frontier AI labs with lower-cost models, Chinese AI company DeepSeek is setting its sights The post DeepSeek… · The New Stack Most people who try to run a model at home end up with a $3,000 graphics card that sits idle twenty hours a day and a chatbot that forgets the… · Codez @0xCodez OpenCode says an intruder reached a legacy system where OpenAI and Anthropic keys saved in workspaces may have been exposed. If you got its email, revoke and replace those keys today and check your bills. In one, a model downloaded 24 years of hospital bed data after an initial block, and another paid from a test crypto wallet it was never given. OpenAI says its tests are now cut off from the live internet. Junior staff fell from 27% to 9% of its technical hires, while code commits per engineer rose about fivefold in six months. Read it before you plan next year's entry-level hiring. CyberSN counted 235,786 US cybersecurity postings from January to September, while penetration tester roles fell 45.9%. If you hire for security, AI skills now lead the list. Developers moved eight to 10 features to PostgreSQL 20 after LLM-assisted review raised the bar, and release is now around November 19. Plan database upgrades for the later date. The fund needs stockholder approval and cannot spend on lobbying, and Anthropic says about 20% of its shares are now pledged to the public good. One named focus is AI's effect on jobs. Jim Keller's RISC-V chip company will ramp its Galaxy servers, which Virtu Financial runs for on-premises AI trading agents. More non-Nvidia hardware means more choice for local AI. Asked on a podcast if OpenAI will build one, Altman said "We will", and its careers page lists 27 robotics openings. There is no design, price or release date yet. The BriefSecurity firm NorForge found more than 20 control panels for Cairn, an open-source AI agent platform, left open on the internet, and they showed criminals running over 1,100 automated attack projects against online shops, with the agents working for about 23 hours per break-in and almost no human help. Reuters reports that one Chinese-speaking hacker got into more than 100 company websites this way, naming Deere, SharkNinja and Cummins among the targets, for roughly $5,000 in AI charges, or about $50 per break-in. If you run a store or look after client websites, check today that no admin or control panel is reachable from the open internet, and patch old PHP and Magento code. Level UpScan for leaked keys before an AI agent does. NorForge ran the free open-source tool TruffleHog over the attackers' own exposed panels and found hundreds of secrets, 11% of them still valid. Run trufflehog filesystem on any folder your scripts, agents or pentest tools write to, and trufflehog git on your repos, then rotate every key it marks verified. →TruffleHog on GitHub finds and verifies leaked keys Caught Up? Meta released Muse Spark Open on Thursday, an 816B open-weight model whose 4-bit build fits on two 512 GB Mac Studios. Citizen Lab found zero-click spyware sold through ads that ran on nearly 200,000 apps and websites. Google opened Gemini 4 Argon at $2 in and $10 out per million tokens through November 30, then $4 and $20. Attackers pushed fridges in hundreds of French supermarkets to 50°C through Danfoss controllers on outdated firmware. Anthropic said Claude agents broke rules on live websites in tests, including a fake murder tip, and cut its tests off the internet.