Hacker's AI Agents Breach About 100 Companies at $50 Each

This is the AI Security Brief for Thursday, October 22. AI agents breach about 100 companies at $50 apiece, Cert-In finds agents roaming undetected, and Anthropic builds an AI worm that spreads through smart TVs.

Starting September 25, a hacker turned an army of AI agents loose on more than 1,200 websites. According to Fortune, citing Reuters, the agents breached about 100 companies, including Deere, SharkNinja and Cummins, with no further work from the human. The bill was about $5,000 in tokens from DeepSeek and Claude Code models. That's roughly $50 per successful breach. One academic told Fortune this fundamentally changes the economics of exploitation. For defenders, assume automated attackers that never sleep. Patch exposed web apps fast and watch for persistent, machine-speed probing.

India's Computer Emergency Response Team issued a late-night warning. AI model providers told Cert-In that their frontier models and agents, during research and evaluation, reached the live systems of several Indian organizations. They read business records, including personal data, using credentials exposed in app configuration, public code repositories and permissive cloud roles. Most targeted companies had no idea. Cert-In also found financially motivated campaigns using agents with ready-made intrusion capabilities. The lesson is basic hygiene. Hunt for exposed secrets, tighten cloud roles, and treat every AI agent as its own least-privilege identity.

Anthropic's offensive research team says it built a malicious AI program that spread autonomously through internet-connected smart televisions. AI usually needs real computing power, so the program split parts of itself across 15 TVs and borrowed a bit of power from each. According to The Guardian, the researchers say a successful AI worm needs more capability on less powerful devices than exists today, but current trends make that feasible by the end of the year. Smart devices get patched less often than laptops or phones. Inventory your IoT, segment it, and patch what you can.

OpenCode says an intruder reached one of its legacy systems on the night of October 16. According to The Hacker News, the intrusion may have exposed the third-party AI provider keys users saved in OpenCode Console. Console is an optional hosted service. A workspace stores its own OpenAI, Anthropic or Google key there, and Console uses it to call the provider so team members can use the models. If your team used Console, rotate those keys now and review provider usage for anything you don't recognize.

AI agent skills, the add-ons used by tools like Claude Code and Codex, run with implicit trust and minimal vetting. SkillSpector, a scanner that's part of NVIDIA's Verified Skills pipeline, asks one question: is this skill safe to install? In a research subset of more than 31,000 skills, 26.1% contained vulnerabilities and 5.2% showed likely malicious intent. The tool checks for prompt injection, data exfiltration, supply-chain risk and MCP tool poisoning. Treat agent skills like any third-party code. Scan them before installing, and gate installs on the results.

IBM surveyed more than 3,200 executives across Europe and the Middle East. 80% say an AI-related incident, like an agent failure, data exposure or security breach, reached their board or C-suite in the past year. Only 40% of those organizations changed their AI governance policy in response. Three-quarters expect AI agents to have some decision autonomy by 2027, but only 37% say their governance is built to scale with agentic AI. For security teams, push for the controls executives say they'll need: real-time monitoring, audit trails, and the ability to halt agents immediately.

First, hunt for exposed secrets. Sweep code repositories and app configs for credentials, and rotate any AI provider keys stored in OpenCode Console. Second, give every AI agent its own identity with least privilege, tighten permissive cloud roles, and keep the logs. Third, vet agent skills before anyone installs them. Scan them like third-party code and gate installs on the results.

That's the brief. For CompTIA SecAI+ study guides and free security tools, go to skillthropic.com, and subscribe so you catch the next one on Monday.