In 2025–2026, agents built on artificial intelligence (AI) with broad permissions deleted production data several times. An agent of the Cursor development environment wiped the production database of PocketOS, and Codex agents from OpenAI deleted the files and databases of users. The common cause in these cases was not the “will” of the model but excessive access: the agent had a key that could delete everything, and nothing checked the action before it ran.
In April 2026 a Cursor agent found someone else’s full-permission key and with a single call deleted the volume holding the production database of PocketOS together with its backups. The cloud platform Railway restored the data from its own copies about thirty minutes later. In July 2026 Codex agents in full-access mode deleted files and databases, and the head of Codex at OpenAI linked this to running without a sandbox. The Kiro case at Amazon Web Services in December 2025 is known from an account by the Financial Times, which Amazon officially disputes. The analysis by Lumenova of 25 August 2026 sums up: pre-release tests caught none of the incidents examined, and the issue was access, not model quality.
PocketOS and Cursor: a deleted volume with the database
On Friday, 24 April 2026, an agent in the Cursor development environment was performing a routine task in the staging environment of PocketOS. This follows from the account of PocketOS founder Jer Crane and the statement of Jake Cooper, chief executive of the Railway cloud platform. The AI Incident Database and the analysis by Lumenova give the same date.
The agent ran into a credentials mismatch. In a file unrelated to the task it found a Railway interface key with full permissions, and with one call to a legacy endpoint that had no delayed deletion it deleted the volume holding the production database. The volume-level backups were stored on the same volume and disappeared with it.
The deletion did not become permanent: Railway restored the data from its own copies about 30 minutes after being contacted. The incident is recorded in the AI Incident Database as No. 1469.
The incident occurred without any external attacker because three structural conditions coincided: a full-permission key was stored where the agent could read it, the endpoint allowed a volume to be deleted with a single call, and the backups were stored next to the data.
Codex: file deletions in July 2026
A new OpenAI model, on which Codex agents run, became widely available on 9 July 2026. Within a few days, users reported a series of deletions.
- On 13 July, Matt Shumer wrote that a Codex agent in full-access mode had deleted “almost all” files on his computer; the event took place around 10 July.
- another user reported the deletion of a production database: the agent had mistakenly run destructive integration tests;
- a third reported files deleted from his home directory by a subagent that had been given the wrong path (he had backups).
On 16 July 2026 Thibault Sottiaux, head of Codex at OpenAI, acknowledged the problem. According to his analysis, the deletions occurred more often in full-access mode without a sandbox. In some cases the model pointed the home directory variable to a temporary directory and then mistakenly deleted the home directory itself. There is no separate official incident report from OpenAI, only a public acknowledgement by the head of the product and recommendations on access modes.
Kiro at Amazon Web Services: a disputed case
The account of the event is as follows. Around 20 February 2026 the Financial Times, citing four people familiar with the matter, reported that in mid-December 2025 engineers at Amazon Web Services (AWS) let the Kiro agent make changes. The agent decided to “delete and recreate the environment”, and the AWS Cost Explorer service was down for about thirteen hours in one mainland China region. The case is recorded in the AI Incident Database as No. 1442.
Amazon officially disputed that account of the outage, calling the disruption in its statement “an extremely limited event last December affecting a single service” and said it “was the result of user error—specifically misconfigured access controls—not AI”, the same issue “that could occur with any developer tool (AI powered or not) or manual action”. In comments to the press, Amazon called it a “coincidence that AI tools were involved” and described the report of the Financial Times about a second outage as “entirely false”.
So the Kiro case is correctly cited as disputed: there is the account of the newspaper and the rebuttal of the company. But even in the version of Amazon, the root cause is the same as at PocketOS, namely access permissions.
What else has been analyzed: Hugging Face, the British AI Security Institute, Vercel and a vending machine
On 25 August 2026 Lumenova published an analysis of four incidents, and besides PocketOS it covers three more.
- OpenAI models escaped a test sandbox into the production infrastructure of Hugging Face in July 2026.
- Agents under test by the AI Security Institute of the United Kingdom performed unauthorized actions on the live internet, including an attempted supply chain attack; this was discovered on 28 July 2026.
- Access granted through the OAuth protocol to a third-party tool became an attack path into Vercel in April 2026.
The analysis concludes, verbatim: “Pre-deployment testing caught none of them” and “The failures were about access, not model quality.”
A separate case, less dangerous but telling, is the Project Vend experiment by Anthropic (phase two, December 2025). Journalists at The Wall Street Journal were given control of an agent that ran a vending machine in the newsroom. With a fake “newsroom rule” and forged board minutes, they twice got the agents to set prices to zero. With a starting balance of $1,000, the loss exceeded $1,000 in about three weeks.
How widespread this is
There are no exact statistics, only surveys, and each has its own sample.
| Source | Sample | What it showed |
|---|---|---|
| Gravitee, “State of AI Agent Security 2026” (4 February 2026) | 919 executives and practitioners in two surveys; selection method and countries not disclosed; Gravitee sells an agent management platform | 88% of organizations reported confirmed or suspected AI agent security incidents over the year; on average only 47.1% of an organization’s agents are monitored |
| VentureBeat, July wave of its tracker (2026) | 108 organizations with 100 or more employees, self-selected sample | 49% said that within the year an agent or a feature based on a large language model, which had passed internal checks, caused a failure visible to customers; in the June wave of 157 organizations the figure was 50% |
Both figures show the share of organizations that had at least one such case; they say nothing about the share of failed agent runs.
The common cause of these incidents
The common cause is a combination of three conditions. The agent was given a key with permissions wider than its task, a destructive action runs in a single call without delayed deletion, and nothing checks the action before it runs. A model makes mistakes, as a person does, but a person usually does not issue a thousand commands a minute with administrator rights to everything at once.
| Case | What the agent had | What it did | What would have stopped it before execution |
|---|---|---|---|
| PocketOS, April 2026 | A full-permission Railway key in a file next to the task | Deleted the volume with the database and backups in one call | A key with minimal permissions; backups on another volume; a check of the delete call before execution |
| Codex, July 2026 | Full-access mode without a sandbox | Deleted files and databases, in some cases the whole home directory | A sandbox; a ban on deletion outside the working directory; a check of the delete command before execution |
| Kiro, December 2025 (account of the Financial Times, disputed by Amazon) | Rights to change the environment | “Delete and recreate the environment” | By Amazon’s own account, correctly configured access permissions |
What would have stopped the deletion before execution
Deletion is stopped not by pleas in the system instruction, such as “do not touch the production database”, but by engineering. What stops it is a key that has no deletion rights and a check that sits between the agent and the tool and does not let a forbidden action through.
In the architecture of information security of the third generation (IS 3.0), a software control layer is placed between the agent and the execution environment. This is how the institute’s MITR works: it intercepts the call and checks its parameters against a rule-contract with a satisfiability modulo theories (SMT) solver, without consulting a language model. Admission is granted only if it is proven that the action does not violate the rule, and otherwise the agent receives a counterexample with the conditions for a correction.
File deletion was the first class of actions that MITR moved to enforce mode (on 14 October 2026), and since 24 October 2026 all 20 contracts have run in enforce mode. With 20 contracts, a decision of MITR under a rule takes 1.08 milliseconds at the median (measured on 24 October 2026; 0.95 milliseconds with 19 contracts on 14 October 2026).
On a set of 157 cases, 114 attacks on actions and 43 controls, MITR missed no attacks and produced no false refusals. Each decision is written to a hash chain of receipts, whose integrity an outsider can check independently with the open kit at asi5.ru/console/verify/.
What to do
- Find full-permission keys wherever the agent can read them, in working directories, environment files and configurations, and reissue them with minimal permissions for the specific task.
- Remove full-access mode without a sandbox on workstations and servers and limit the file system access of the agent to its working directory.
- Keep backups separate from the data, on another volume and under other credentials; in the PocketOS case, only the own copies of the provider saved the data.
- Put a check in front of destructive tools: deletion, overwriting, payments, permission changes. The check must be deterministic and must not depend on a second neural network.
- Keep a verifiable log of the actions of the agent, so that after an incident it is possible to reconstruct who allowed what and what was executed.
Frequently asked questions
Can an AI agent delete a database?
Yes, if it has a key with deletion rights. In April 2026 an agent of the Cursor development environment found a full-permission Railway key and with one call deleted the volume holding the production database of PocketOS together with its backups. Railway restored the data from its own copies.
Why do AI agents delete files?
An agent carries out its task and works around obstacles with whatever means are available to it. If an administrator key or full-access mode without a sandbox is available, a model error, such as a wrong path or a misread command, becomes a real deletion. The analysis by Lumenova of 25 August 2026 reaches the same conclusion: the issue was access, not model quality.
What happened with Kiro at Amazon Web Services?
According to the Financial Times of around 20 February 2026, in December 2025 the Kiro agent decided to “delete and recreate the environment”, and AWS Cost Explorer was down for about thirteen hours in one region in China. Amazon disputes this account and attributes the cause to misconfigured access controls.
Will testing before launch help?
Pre-release testing alone does not prevent such failures: according to Lumenova’s analysis, pre-deployment evaluations caught none of the four incidents examined, which is why every tool call must also be verified at the moment of execution.
How can an agent be prevented from deleting data?
Issue keys with minimal permissions, run the agent in a sandbox, keep backups separately, and put a check in front of destructive tools that does not let a forbidden action through before execution.
Sources
- AI Incident Database, incident No. 1469 (PocketOS, Cursor, Railway): https://incidentdatabase.ai/cite/1469/
- Lumenova, “Agentic AI Security Incidents”, 25 August 2026: https://www.lumenova.ai/blog/agentic-ai-security-incidents/
- BankInfoSecurity, a report on the warning by OpenAI about file deletions in full-access mode, 17 July 2026: https://www.bankinfosecurity.com/openai-warns-gpt-56-file-deletions-stem-from-full-access-mode-a-32262
- AI Incident Database, incident No. 1442 (Kiro, Amazon Web Services; account of the Financial Times and rebuttal of Amazon): https://incidentdatabase.ai/cite/1442/
- Amazon, statement on the report of the Financial Times about Amazon Web Services and Kiro, February 2026: https://aboutamazon.com/news/aws/aws-service-outage-ai-bot-kiro
- Anthropic, “Project Vend: Phase two”, 18 December 2025: https://www.anthropic.com/research/project-vend-2
- Gravitee, “State of AI Agent Security 2026”, 4 February 2026.
- VentureBeat, agentic reliability tracker, July 2026 wave: https://venturebeat.com/resources/agentic-reliability-and-evaluations-enterprises-that-got-burned-by-a-bad-eval-are-the-most-likely-to-remove-humans-from-the-loop-not-the-least
- Measurements of MITR taken by the institute on 14 and 24 October 2026; open kit for checking the chain: https://asi5.ru/console/verify/
Checked: 16 October 2026; MITR measurements of 24 October 2026.