An Authorized Task Is Not an Authorized Retry
A financial-services firm gave an agent write access to its billing database and a legitimate job: reconcile ledger anomalies. One corrupted value broke a tool, and in under an hour the agent's attempts to fix it generated more than fifteen thousand reasoning calls. Nothing in the account points to a stolen credential or a revoked mandate. What the initial authority to do the task could not answer, at any point in those fifteen thousand calls, was whether the next attempt should still be allowed to run.
What happens
Mandiant’s September 2026 AI Risk and Resilience report describes an incident at a global enterprise financial-services provider. The firm had deployed an agent to reconcile anomalies in its accounting ledgers and, to let it do the job, granted it direct read and write access to internal billing databases. This was not an experiment at the edge of the organisation; it was an agent acting on production financial records under a mandate it had every right to hold.
Then a corrupted null value broke the agent’s formatting tool. Faced with a failure it could not parse, the agent did what an unconstrained system does with a problem: it tried again. In Mandiant’s words, it “entered an unconstrained, recursive reasoning loop to brute force a fix.” In under an hour it generated more than 15,000 high-frequency, high-cost reasoning API calls, triggered a cloud-billing spike of roughly $50,000, and caused database locking severe enough to halt live business transactions.
Mandiant’s recommended controls are specific: identity and role-based access for agents, cost-cap thresholds, and — the operative one — “automated financial circuit breakers designed to halt agent operations after a set threshold of consecutive task failures,” alongside bounded recursion limits and rate limits at the service-identity and project level. The report files the incident under governance, and calls the failure mode a runaway execution loop.
The usual reading
Read as an operations story, this is a cost-control failure: operational controls that did not sufficiently restrain repeated execution and rising cost, on an agent billed by the call. Read as an AI story, it is a reasoning failure: a model that could not recognise a dead end and kept pushing.
Both readings are true, and both locate the problem inside the agent or the bill. Neither asks what the agent was authorized to do at call fifteen thousand — and whether that was the same thing it was authorized to do at call one.
What the report establishes, and what it does not
Start with what the report establishes. The task was legitimate when it began. The agent was granted the access it used. Nothing in the account describes a stolen credential, a hijacked session, or a mandate that was revoked and ignored — this is not an intrusion case, and the report does not file it as one. What the report does not provide is a record of the agent’s authority over the course of the hour: whether any grant carried retry, time, or cost conditions, or whether anything about its permissions changed. On the authority dimension, then, the honest statement is narrower than “the agent was permitted throughout”: the account shows repeated execution producing harm without any change in authority being part of the story.
Now what the report does show changing. The formatting tool was broken — the agent’s own capability to complete the task had failed. The database ended up locked severely enough to halt live transactions unrelated to reconciliation. The cost of the work reached a five-figure spike. The report gives these as outcomes, not as a timeline; it does not record at which attempt the loop should have stopped, and this piece does not claim to. Each of these is a fact about the current state and current conditions of the execution, and each of them bears on whether continuing was still eligible — independently of whether the agent was still permitted. What the outcome establishes is that this repeated execution was not constrained before it reached the cost spike and the interruption of live transactions.
That is the distinction this case makes visible. Authority granted to a task — and such grants can carry conditions of their own on scope, time, or cost — is one thing. Whether the conditions for the current attempt are met is another. The report gives no reason to think the first was withdrawn. It gives every reason to think the second had stopped holding by the time the loop did — and no record from which to say how much earlier.
Every retry is an execution
The mechanism matters, because it is where the boundary would sit. A retry is not a continuation of a decision already made. It is a new attempt to act — a new call, a new draw on a metered API and, where it writes, on a shared database — made under conditions that the previous attempts have altered. The report gives the total, more than 15,000 reasoning calls, not a sequence; it does not say how many of those calls wrote to the database or how many failed. But the outcome makes the structure clear: the cumulative cost, the state of the target database, and whether the tool had been repaired are conditions a next attempt would depend on — and, on this series’ reading, conditions that should be established before a next attempt opens, whatever the loop actually consulted. This is not a claim that a person should re-approve each attempt. It is a claim that before a new attempt consumes resources or changes state, something should establish that the conditions for continuing still hold.
This series has looked at the gap between an approval and the execution it enables from several directions: a mandate that stays valid while the world moves, a deferred purchase whose consent is history by the time stock returns, a corrected state that drifts. This case is the same gap, folded inward. The conditions that matter can change not only between an initial approval and its execution, but between one execution and the next — and the report shows that gap is as real at a second’s remove as at a week’s.
Mandiant’s proposed control describes this: halt “after a set threshold of consecutive task failures,” bounded by recursion limits and cost caps. Read structurally, that is an eligibility control on stated conditions — the failure count, the cost, the recursion depth — applied before the next attempt, and it does stop the next execution when those conditions are exceeded. It does not question whether the agent may reconcile ledgers. It questions whether this attempt, now, is still within what was meant. The question this series adds is what such a control is looking at: whether the conditions it consults reach the current state of the target, the recoverability of what has already been done, and the effect of the next action — or stop at those limits.
What this does not show
The account is Mandiant’s, anonymised, and not independently verified; the figures — 15,000 calls, ~$50,000, under an hour — are the report’s. The report does not describe what controls the firm had in place before the incident beyond the access grant, whether any rate limiting or budget alerting existed and failed, or how the loop was eventually stopped. This is not a finding that the firm had no execution controls; it is a finding, on the report’s own terms, that repeated execution was not sufficiently restrained before it reached the reported harm.
Nor is the lesson that agents should not be given write access to production databases; the report describes the grant without ruling on whether the design required it. The lesson is narrower: the fact of an initial grant does not by itself establish that the conditions for the next attempt are met, and a system that checks only the grant will let an authorized actor continue after continuing has stopped being eligible.
Where this is heading
Most of the execution-boundary discussion this year has concerned the moment before an action opens — whether the mandate is current, whether the scope matches, whether the state is what the approval assumed. This incident is a reminder that the boundary is not a single door. An action that has been permitted to open can still be re-evaluated while it runs, and for anything that retries, loops, or continues — reconciliation agents, remediation loops, long-horizon workflows — the questions are the same ones asked at the door: is the authority still current, is the state still the one the task assumed, are the conditions still within what the authorization was meant to cover, and if they cannot be established, does the next attempt hold rather than fire.
The controls Mandiant recommends — failure thresholds, recursion bounds, cost caps — are a real form of that re-evaluation: they decide, on stated conditions, that the next attempt may not open. The finer form widens what is consulted — beyond failure, cost, and recursion limits to the current state of the target and the recoverability of what has already been done.
The proposition
A task can be authorized once and executed many times. Each execution opens under conditions the previous one has changed: the state of the target, the cost incurred, the count of failures, the capability of the agent itself. Authority granted to the task is one object, and it may carry conditions of its own. Whether the conditions for this attempt are met is another. When the second is not consulted before each attempt, the first will carry an agent from a legitimate job to a locked database and a fifty-thousand-dollar bill with no change in its authority anywhere in the story.
Permission to start is not permission to continue. Every retry is a new execution, and the world has moved since the last one.
Source: Mandiant, AI Risk and Resilience Report 2026, “Case study 6: ‘Denial-of-Wallet’ using rogue reasoning loop” (Google Cloud, September 2026). The incident account, figures, and recommended controls are the report’s; the organisation is not named; the reading is this series’.