AI agents are deleting real data
The short answer. In 2026, AI agents from several vendors ran commands on real machines and destroyed real data. The checks that failed mostly lived inside the agent, so the limit has to be enforced from outside it, and recorded where the agent can't reach.
What are agents doing on real machines?
They run shell commands with your rights, so one wrong command can wipe whatever your account can reach. The clearest case is a delete whose path collapses. On Windows, nested quoting in cmd /c rmdir /s /q can turn a folder path into a drive root. Cursor forum threads from April to September 2026 report about 300 GB lost, a reinstalled Windows and most of a D: drive. Cursor staff gave the cause in the first: "on Windows we don’t have a filesystem sandbox (like seatbelt on macOS or Landlock on Linux)." Cline #14864 (6 October 2026) reports the same rmdir /s /q wiping about 1.5 TB of photos and documents from a data drive.
Claude Code #88462 (21 August 2026) reports rm -rf on a WSL2 home directory and lists five earlier reports like it, so it isn't only Windows. For Codex, Simon Willison quoted OpenAI's Thibault Sottiaux: GPT-5.6 deleted files most often when "Full access mode is enabled and codex is run without sandboxing protections", after trying to override $HOME for a temporary directory.
In Codex #59060 (3 November 2026), a user asked the agent to reboot a Linux server. The agent later wrote that it "wrongly treated the problems I saw afterward as permission to edit /etc/fstab and start containers." The reporter puts the cost at over $10,000. A commenter: "If you need AI to reboot a server you maybe shouldn't have the permissions that the AI had in order to do what it did."
Why didn't the safety checks stop it?
They sat inside the agent, where its own mistake could reach them.
- A string check sees a string. In #88462 the agent wrote a script with a delayed
rm -rfon a variable it later set to$HOME; the permission check saw onlybash /tmp/.../test-lib-demo.sh. - The model can waive the approval. Cline #12020 (1 July 2026) reports the gate trusts the model's own
requires_approvalflag, so a destructive command markedfalseruns without asking. - An attacker can switch approvals off. The 2025 Nx malware ran AI CLIs with flags such as
--dangerously-skip-permissionsand--yolo, according to Wiz. Check Point Research found two repository settings that "allow automatic approval of MCP servers"; Anthropic has fixed it. - The guard didn't exist. In Claude Code #34327 the agent said it had added a git hook to block
git reset --hard. The reporter says it was never written, and the reset happened again the next day. - A label isn't a limit. In #75861 a read-only
Exploresub-agent ranrm -rf .claude/worktrees; the reporter says read-only appears to be a prompt, not a tool limit.
Cursor staff also wrote in July that File-Deletion Protection covers the editor's own tools, but "shell commands like rm -rf bypass that protection."
Reporters ask for a guard outside the model. #99198 wants a guard "applied in all modes including bypassPermissions and for sub-agents". Codex #33624 wants one even under Full Access, because "the larger safety property should not depend on one model correctly forming one shell command." #95426 asks the tool to refuse, not prompt, when a delete resolves to a drive root.
Why is agent tooling an attack surface?
Agents load code and settings that anyone can publish.
- Skills. Snyk scanned 3,984 agent skills and confirmed 76 malicious payloads (5 February 2026). Koi flagged 341 of 2,857 ClawHub skills, 335 from one campaign; publishing then needed only a week-old GitHub account, per The Hacker News.
- Config files. The ChainDrop worm committed a
.claude/settings.jsonSessionStart hook and a.vscode/tasks.jsonfolder-open task, StepSecurity found. SANDWORM_MODE injected its own MCP server "into every AI coding assistant config it finds on disk", Socket reported. Check Point showed a committed hook could run commands "on every collaborator's machine". - AI dependencies. Compromised LiteLLM releases shipped in March (1.82.7 and 1.82.8) and in October (1.107.0); the October stealer took login files for AI development tools, according to BleepingComputer.
- Publish tokens. An unauthorized
cline@2.3.0, published with a leftover npm token, installed OpenClaw on install and was live for about eight hours, Cline's post-mortem says. Cline says no malicious code was delivered, and OpenClaw is a legitimate project.
Why does the evidence disappear?
The record sits where the damage lands. In #95426, ~/.claude/projects was deleted too: "The audit trail was inside the blast radius." In #10077 the log kept the tool output but not the command. In Codex #42875, shell history and transcripts were "destroyed by the deletion itself", so the cause couldn't be settled. A useful record has to live where the agent can't write.
What do the surveys say?
Definitions differ, and most come from vendors in this space.
- Gartner: of 360 IT application leaders (May and June 2025), 74% see AI agents as a potential new attack vector, and only 13% strongly agreed they had the right governance (release).
- SailPoint with Dimensional Research (May 2025): of 353 respondents, 80% say their agents took unintended actions; 44% have policies to secure them (release).
- Gravitee: the February 2026 wave (over 900 respondents) found 88% with a confirmed or suspected agent security incident in a year; the June 2026 wave (750 leaders, April fieldwork) found 54%, 34.9% confirmed. The waves aren't comparable.
- Dynatrace (January 2026): of 919 senior leaders, 52% name security, privacy or compliance as a top barrier to production (release).
What are standards bodies and buyers asking for?
The OWASP Top 10 for Agentic Applications (9 December 2025) names Tool Misuse and Exploitation (ASI02), Identity and Privilege Abuse (ASI03) and Agentic Supply Chain Vulnerabilities (ASI04) in its document. Its advice is "to avoid unnecessary autonomy", and it calls strong observability "non-negotiable".
On 1 May 2026, CISA, the NSA and the cyber agencies of Australia, Canada, New Zealand and the UK published joint guidance, Careful Adoption of Agentic Artificial Intelligence (AI) Services. It calls for "ensuring agents are not granted broad or unrestricted access", plus layered defences, oversight and monitoring. NIST's AI Agent Standards Initiative, started in February 2026, aims at voluntary guidelines.
Money is moving too. Zenity raised a $125M Series C in August 2026; HiddenLayer raised a $100M Series B in September, and its release mentions "Agentic Runtime Security capabilities". SailPoint closed its purchase of Entro on 29 June 2026, undisclosed but reported at about $200M. Palo Alto Networks closed its purchase of Koi on 14 April 2026; announcing it, its chief product and technology officer said "AI agents and tools are the ultimate insiders." That shows investors care about agent security; it says nothing about demand for any one product.
Where do existing answers stop?
Prompts and instruction files hold only while the model follows them. Vendor permission modes help, but several of these incidents ran in modes that switch them off (#99198 ran in bypassPermissions).
Vendor sandboxes help, but Cursor's covers only Cursor's agent. NVIDIA's OpenShell runs each agent in a fleet in an isolated container, pod or VM, depending on the driver. Containers isolate well, yet most reports above describe an agent working directly on the user's own files.
What's missing is a host-side control: one that works for any agent process from outside it, with rules the operating system enforces, a person's approval before irreversible moves, and a record the agent can't edit. Some harm even that can't stop. Data can leave through any endpoint the agent may reach, such as github.com (Tego's known limits, item A1), and effects inside a SaaS or a vendor's cloud happen where nothing on your host runs (A5).
Where does Tego fit?
Tego is our open-source host-side control for Linux. It runs an unmodified agent under a contract you write. At every level, the kernel refuses writes and programs outside the contract (README); Basic needs no root, Plus and Max need sudo (protection levels).
An ask rule holds a file action or a connection until a person answers; commands can be refused but not held (README). Every run writes a record that shows any later edit. Tego never calls a model. It has been public since 2 November 2026, so this post argues the need, not demand for Tego.
Our teardowns take these incidents one at a time on the blog.
Try it: Tego on GitHub and the Tego docs. Running agents across a team? Talk to us.
All sources were checked on 5 November 2026.
Sources
Sources checked on 5 November 2026.
Every link in this post, by site:
- github.com/Attention-Span-AI/Tego
- forum.cursor.com
- t/cursor-wiped-300gb-from-my-system-after-a-simple-cache-cleanup-request-critical-issue/159065
- t/severe-data-loss-agents-mis-quoted-rmdir-command-recursively-deleted-my-entire-c-drive-had-to-reinstall-windows/163626
- t/windows-agent-rmdir-broken-quoting-wiped-d-ticket-t-f90960-grok-4-6-extra-high/172345
- t/cursor-claude-code-deleted-my-entire-documents-folder-on-macos/167086
- simonwillison.net
- research.checkpoint.com
- thehackernews.com
- stepsecurity.io
- bleepingcomputer.com
- sailpoint.com
- gravitee.io
- dynatrace.com
- content.govdelivery.com
- hiddenlayer.com
- paloaltonetworks.com
- docs.nvidia.com
Read next
Eleven agent incidents and the rule for each
Eleven agent incidents, each in one line.
Asked to reboot, the agent edited /etc/fstab
Strict rules refuse key steps
Your agent can ignore CLAUDE.md
A rule your agent can read isn't a rule it must follow.