
Table of Contents
THE SIGNAL
Agentic attack campaigns get described in terms of speed. The number I watch is cost per attempt, and this week we found out who is paying it.
Amazon's MadPot decoy network captured attacker-run AI agents, assessed with moderate confidence as ARTEX-based and operated by separate China-linked groups, sweeping the internet for self-hosted AI applications: LiteLLM, the One API family, Dify, FastGPT, Langflow, RAGFlow, Open WebUI, LobeChat and Ollama. The target was the AI provider keys stored inside them. Each valid key went into a shared pool. The playbook reads like an audit finding list: LiteLLM's default master key, management endpoints with no login, a header-based auth bypass in One API, SSRF to cloud metadata, secrets left in JavaScript, source maps and env files. When the agents found cloud credentials, they went for managed AI model services first and general cloud access second. Last week Amazon described the economics plainly: stolen keys paid for the LLM inference that powered the next round of credential hunting.
The same week showed where those keys come from. Cursor confirmed that a key exposed at its inference provider Fireworks AI let an attacker copy historical coding-agent traces from a shared bucket. Some of those traces held unredacted API credentials. Those led into Cursor's internal SaaS tools and eventually to a credential that put trojanized Claude Code extension builds (2.1.312 and 2.1.313) into its marketplace cache for about 11 hours, reaching up to 5,000 developers with an SSH backdoor. The same reporting puts Railway's bill at roughly $740,000 in fraudulent AI charges run up on stolen customer secrets. Barracuda disclosed that an attacker used ARTEX and a cloud key issued to an AI vendor for a finished project, never retired and sitting in the vendor's GitHub config, to move through about ten cloud accounts in 44 hours. And for roughly 45 minutes, Microsoft's own agent-framework package on PyPI shipped a stealer that ran on every Python start and requested fresh Azure managed-identity and GitHub Actions OIDC tokens. The package gets about 190,000 downloads a month.
Here is the part I hadn't seen spelled out before. Draw it as a system, a stock with two flows. The stock is live credentials sitting in the AI layer. The inflow is fast: every new integration mints a key, gateways like LiteLLM concentrate the keys for every provider in one box, agent traces log them, vendors receive them. The outflow is expiry, rotation and offboarding, and Barracuda's key shows how slow it is. Attackers have now built a reinforcing loop on top of that stock: keys buy inference, inference runs more agents, agents find more keys.
A reinforcing loop grows until it hits a limit. In most criminal economies the limit is the attacker's budget. This one runs on the victims' budget, so the only limit left is the size of the stock. We own the stock.
Sources: The Hacker News (Amazon MadPot), The Record, BleepingComputer (Cursor), BleepingComputer (Barracuda), The Hacker News (agent-framework).
THE MAP

Their budget is your key hygiene. The attackers' loop draws on a reservoir we fill and barely drain. Patching slows the pump a little. Shortening how long a key stays alive takes away what it pumps.
Product teams aren’t short on ideas. They’re missing a system.
Jira Product Discovery gives teams one place to capture customer feedback, prioritize ideas with consistent frameworks, and build living roadmaps everyone can align on. And when it’s time to build, those decisions connect directly to delivery in Jira, so everyone can see how the roadmap turns into real work.
AROUND THE PERIMETER
VMware vCenter, CVE-2026-123890: Armory Castle (formerly UNC5221) exploited this authentication bypass as a zero-day for more than two months, entering through a BeyondTrust Remote Support zero-day and then forging SAML tokens to get past MFA. There is no fix for end-of-support 7.0, and the exploit tool works against it. If you still run 7.x, isolate it and plan its removal, because no patch is coming.
Optimizely, OPTIMJACK: an attacker wrote a loader into customer projects and Optimizely's own CDN served it to visitors of 178 organisations, HSBC and Equifax among them: a card skimmer, a crypto address swapper, and iOS exploit kits. A trusted origin passes every allow-list, so check which third-party tags load on your login and payment pages.
LiteLLM, CVE-2026-40217: now in CISA's KEV catalog, exploited to turn limited inference keys into full admin credentials in LLMjacking campaigns against financial services and others. This is the Signal arriving through a CVE number.
Cisco Expressway, CVE-2026-76466: hard-coded credentials give unauthenticated root on clustered devices, and it was exploited shortly after the PoC went public. Patch to X15.5.20 and block TCP 4371–4372 except between cluster nodes.
Puppet, CVE-2026-115731: exploited RCE on Puppet Server, reachable without credentials where certificate auto-signing is loose. Your configuration manager can push to every host, so treat it as tier zero.
CALM THE NOISE
The loudest stories this week were "the AI found a zero-day" and Google's forecast that fully autonomous attacks are three to six months away. Both deserve a closer read than they got.
Cairn's zero-day is real. Amazon reports agents going from mapping the API to exploitation in 11 minutes. But the target was MONAI Label, a medical-imaging labelling server that by default runs with no login, and the flaw was path traversal. In the same campaign the agents spent 16 hours and 148 actions failing to take over a Redis server, then tried more than 30,000 credential combinations against a GPU server over six hours and got nowhere. The operator's own Cairn control panel sat on the internet without a password. Google found exploit_agent, its first fully autonomous patch-to-exploit case, on a staging server that was also exposed. For all the talk of machine-speed offence, the operators run their own infrastructure about as carelessly as their victims do.
Google also notes a flood of AI-generated PoCs is adding to the verification burden. NGINX2Root shows why: the public PoC needs about 128 GB of HTTP/2 traffic. Patch exposed HTTP/2 instances, and don't move it ahead of the five exploited flaws above.
The lab stories, California suing OpenAI over 80-plus loss-of-control incidents and Gemini agents hammering Epoch AI's live site with 36,411 requests, matter for governance. One lesson carries over: those agents were told "NEVER stop, NEVER give up" and had no stop condition. Check that your own agents have one. Then go back to your keys.

