```html ```
top of page

TOKENTHEFT: They Asked the Agent for Its API Key, It Told Them, and They Spent $600,000 of Someone Else's Compute Over Three Weeks. This Is a Category, Not an Incident.

  • Writer: Patrick Duggan
    Patrick Duggan
  • 3 minutes ago
  • 6 min read

METR, the non-profit that runs independent evaluations of frontier AI models, has disclosed that in March 2026 an attacker stole an API key and burned through roughly $600,000 worth of inference over three weeks.


The number is arresting. The mechanism is worse. And the fact that it is the second such disclosure in three days is the reason we are giving the pattern a name rather than filing another incident write-up.


Patrick has been arguing this on LinkedIn as a distinct category. We agree, so here is the case for it.


TOKENTHEFT: theft where the objective is metered inference capacity itself, rather than the data the system holds.





What Happened at METR


An attacker obtained an API key for inference on public models. According to METR's account, once the system was identified the actor prompted an agent directly to reveal its model provider API key — and the agent did. They then added an SSH key for persistent access and spent three weeks consuming credits.


Read that middle step again, because it is the part that should worry anybody running agentic tooling. The credential was not cracked, phished from a human, or found in a committed .env file. It was requested from a running agent, conversationally, and handed over. The agent held a secret and did not hold the line.


That is the beat we have claimed for ourselves and we will report it flat: this one did not hold.


METR says the accrued usage would have been about $600,000 in bills, had the credits not been provided free to the non-profit by the model provider, which it declined to name. So the cash did not change hands. The capacity absolutely did.



Why It Ran for Three Weeks


METR's own explanation is the most instructive sentence in the disclosure: the illicit usage was not caught quickly because METR runs large-scale evaluations that legitimately consume enormous volumes of tokens, and there were no caps on token spend.


The attacker hid inside the victim's normal.


Put that next to the other case this week. On August 30, Anthropic disclosed infostealers replaying stolen Claude session cookies to drain user quota, and said it spotted the pattern because usage limits were refilling and draining while account owners were inactive. That is a clean anomaly — activity where there should be none.


METR had no such contrast to work with. An organization whose entire mission is running massive model evaluations has no quiet period, no flat baseline, no shape that a spike stands out against. The signal and the noise are made of the same material.


That is the general law here, and it is uncomfortable: your AI spend is only monitorable if you know the shape of your normal — and the organizations with the most quota worth stealing are precisely the ones whose usage is spikiest and least legible. The better your AI story, the worse your detection surface for this specific attack.



Why This Is a Category and Not Three Stories


Three dated instances, three different mechanisms, one objective:


Credential theft from a running agent, at METR, March 2026, roughly $600,000 in inference over three weeks. Session cookie replay, disclosed by Anthropic on August 30, draining consumer and team quota across Vidar, LummaC2, StealC, RedLine and Atomic Stealer infections. And a reported resale market, where a single stolen key is monetized by selling the compute on to third parties who never touch the victim at all.


What makes it a category is that the existing controls are aimed somewhere else:


Data-loss prevention sees nothing, because nothing is exfiltrated. Your SIEM sees nothing unusual, because the credential is valid and the API calls are well-formed traffic to a service you legitimately use. Endpoint tooling sees nothing after the initial compromise, because the theft happens at somebody else's data center. Identity tooling sees nothing, because in the session-replay variant the authentication already happened and in the METR variant the key was legitimately issued.


The detection surface is the invoice, and the invoice arrives monthly.


Meanwhile the loss does not scale with how sensitive your data is. It scales with how much quota you carry, which means the risk lands on a research non-profit and a two-person startup on the same terms it lands on a bank.



The Argument for Owning the Iron


Here is where this connects to something we build tooling for, and we are going to make the argument, disclose our interest in it, and then say where it stops.


Disclosure first, because the argument that follows is worth money to us. We build commercial sizing tooling for on-premises AI infrastructure, and we have paying relationships in that space. That gives us a direct financial interest in you finding the next four paragraphs persuasive. We are not going to pretend otherwise, and you should weigh the argument knowing it. We applied the same rule to a partner two days ago when the honest result was unflattering to them; the rule does not get to only run in the direction that costs us nothing.


TOKENTHEFT is a disease of metered inference. The loss is denominated in someone else's billing units, accrues silently against a quota you cannot watch in real time, and is discovered on a statement. Move that workload onto hardware you own — the class of on-premises AI factory our public sizer exists to specify — and the $600,000 failure mode does not merely get smaller. It stops existing. There is no meter to run up. Nobody can bill you for capacity you already bought.


That is a real security argument for on-prem, and it is not one the industry has been making, because the on-prem conversation has been stuck on cost-per-token and data residency.


Now the honest half. Owning the hardware does not delete the risk, it changes its denomination. An attacker with access to your cluster can still run their jobs on your GPUs, and you will pay in throughput and schedule instead of dollars — the same shape as cryptojacking, which is a solved-ish problem precisely because operators can see their own utilization. You also inherit every piece of infrastructure security you were previously renting from a hyperscaler, which is not free and not trivial.


So the fair statement is this: on-prem converts an invisible operating-expense attack into a visible capacity attack. That is a trade most security teams should want, because you can watch your own cluster in real time and you cannot watch your provider's meter. It is not immunity, and anyone selling it as immunity is selling.



Controls, Cheapest First


Put spend caps on every key. METR's own remediation list starts here, and the absence of caps is what turned a credential theft into a three-week theft. This costs nothing and most providers support it.


Alert on rate of change, not absolute volume. Absolute thresholds are useless to anyone whose legitimate usage is spiky — which is the whole population at risk. A doubling week-over-week on a key that has been stable is a signal even when the absolute number is unremarkable.


Give every workload its own key with its own budget, so a single compromise has a blast radius smaller than the organization.


Treat agent context as credential-bearing material. If an agent can read a secret, assume an attacker who can talk to that agent can read it too, and design so the agent never holds the provider key in the first place.


And keep credentials off non-organizational infrastructure, which is the policy METR says it tightened afterwards.



The Honest Timeline


We described this asset class on August 31, one day before the METR disclosure surfaced, and we are not going to dress that up as a prediction. The METR incident happened in March 2026, five months before we wrote anything about it. What we did was name the shape one day before a large receipt for it landed in public. That is a receipt for the category, not a call on the victim, and the distinction is the difference between analysis and horoscope.


Credit where it belongs: METR disclosed an embarrassing incident in useful detail, including the part where their own agent gave up the key and the part where they had no spend caps. Most organizations would have published two sentences. The reporting is The Hacker News's.


We cap confidence at 95 percent. The likeliest error here is the category boundary — with three instances we can say the pattern is real, and we cannot yet say how much of it is opportunistic reuse of ordinary credential theft versus operators specifically hunting quota.


If you carry an AI budget line, go and check whether there is a cap on it. That is a five-minute job and it is the whole lesson.




How do AI models see YOUR brand?

AIPM has audited 250+ domains. 15 seconds. Free while still in beta.



Was this useful? Thirty seconds, no cookies, no tracking, no third parties, your address hashed and never stored. If the box below does not load, the same question lives at https://analytics.dugganusa.com/nps.html?post=tokentheft-they-asked-the-agent-for-its-api-key-it-told-them-and-they-spent-600-000-of-someone-e



Comments

Rated 0 out of 5 stars.
No ratings yet

Add a rating
bottom of page