Brief · Both
Attacks got faster. The fixes didn't.
Cybersecuritynews ran: "Anthropic Warns AI Agents Can Now Automate the Entire Cyber Kill Chain." Cue the Skynet takes.
Read Anthropic's September report. It says public agent frameworks like PentAGI provide scaffolding that "effectively automates each step of the cyber kill chain." "Effectively" is doing real work in that sentence. The report also says "Humans remained in the loop by setting the targets of attacks and reviewing exfiltration." People still pick the victim and cash out. The techniques are mostly familiar: stolen credentials, phishing, unpatched edge devices, XSS. What changed is the economics. Faster, cheaper, and a lower skill bar.
As usual with this stuff, the useful parts are boring. The robots aren't coming to kill us, but they do know how to run the same attacks on a shorter timetable.
Malware that rebuilds itself when you catch it. The report says GTG-20006, an actor whose attribution is consistent with Midnight Blizzard, ran an "AI-assisted workflow that automatically rebuilt and re-deployed their toolkit if it was detected by security products." Per the report, capable actors can "close the loop" faster than defenders can write and ship new detections. My take: signature-only defense was already aging. Now it's aging faster.
Secrets in your apps. GTG-50014, suspected ShinyHunters affiliates, ran a credential-harvesting pipeline on 10 AWS EC2 workers. It downloaded 1.8 million distinct Android APKs, decompiled them, scanned them with TruffleHog, and sent verified hits to a Telegram group. Nothing new about hardcoded secrets. They just got harvested at scale.
Token theft. After breaching a SaaS provider, the same crew pulled a session-store dump with "over 2,100 Azure AD token sets spanning more than 40 corporate tenants in about 34 hours. AI agents performed nearly all of the work." One vendor, forty-plus victims.
AI API keys are production credentials now. The report covers stolen AI keys and fraudulent resellers, and it says to buy AI access only through authorized channels. A stolen key gets the thief free attack compute on your bill.
Opinion: none of this needs new controls. It needs the old ones, done before the attacker's agent finishes its coffee.
What to do: - Don't bet on signatures alone. Add behavioral detection, and alert when security tooling stops updating. The report describes malware built to freeze security updates. - Get secrets out of your apps. Run the same scanners attackers run, TruffleHog included, against your own builds. Rotate anything they find. - Shrink token blast radius. Use short-lived sessions. Know which SaaS vendors hold tokens into your tenant. Have a plan to revoke tokens in bulk. - Treat AI API keys like cloud keys. Use a vault, scope them, set spend alerts, rotate them, and buy only from authorized channels. - Plan for hours, not weeks. The report cites breaches measured in hours, not weeks. If your response cycle is a weekly meeting, that's the gap.