Tegoby Attention SpanBlogBook a demo

Eleven agent incidents and the rule for each

AI agents have deleted production databases, wiped home folders and leaked email, and each time the guard that failed lived inside the agent. Below are eleven public incidents, one line each: what happened, and the rule that would have caught it.

Every incident came before Tego's first release, v0.1.0, published 4 November 2026, so each answer is what Tego would have done, checked against what v0.1.0 ships. Each teardown names the exact rule, the protection level it needs and what it leaves open.

The incidents

  • Sep 2025, an MCP server. A fake Postmark MCP server on npm, which Postmark warned about, added a hidden blind copy to every email it sent. No rule catches it: the copy rode inside the one request to Postmark the task needed. Read the teardown.
  • Dec 2025 to Feb 2026, a CI agent. An issue title could steer Cline's triage agent into a package install on the runner, and a token that was still live later shipped an unauthorized cline@2.3.0. A program list without the package manager refuses the install. Read the teardown.
  • 31 Jan 2026, a 24/7 assistant. One web page could steal an OpenClaw gateway token, switch approvals off and run commands, as depthfirst's write-up shows; a disclosed bug, with no reported victims. A contract that lists no shell tools makes the published proof of concept fail. Read the teardown.
  • 20 Feb 2026, an npm worm. Socket reported SANDWORM_MODE, which writes a rogue MCP server into AI assistants' configs days after install. Write rules that keep agent configs closed refuse the plant, when that part of the worm runs under Tego. Read the teardown.
  • 26 Feb 2026, infrastructure. After an old Terraform state file replaced the current one, an auto-approve command deleted a production stack on AWS. A rule that denies terraform destroy refuses that command before Terraform starts. Read the teardown.
  • 12 and 13 Mar 2026, Claude Code. A developer reported an agent running git reset --hard origin/main at session start two days in a row, destroying 12 unpushed commits. A rule that denies git reset --hard refuses it before git starts. Read the teardown.
  • Apr 2026, Cursor. An agent found a Railway token in an unrelated file, and one curl call deleted a production volume and its backups. The no-prod preset refuses the call, because Railway's API is not on the agent's host list. Read the teardown.
  • Apr 2026, a LiteLLM server. A LiteLLM test endpoint started any command a request named, and the PoeLLM botnet took over more than 3,400 exposed servers. Tight program and host lists refuse the download, the miner and the scans. Read the teardown.
  • 4 Aug 2026, an npm worm. ChainDrop ran a credential stealer under Bun and planted a Claude Code SessionStart hook in repositories. A program list without Bun and a host list without GitHub keep the stealer from running. Read the teardown.
  • 21 Aug 2026, Claude Code. A cleanup trap in a script the agent had written ran rm -rf on a user's WSL2 home folder. Letting the agent write only its project folder refuses every delete outside it. Read the teardown.
  • 3 Nov 2026, an ops agent. Asked to reboot one server, an agent edited /etc/fstab and started containers, by its own account. A contract that leaves only the team's gated tool refuses the ssh, aws and curl paths the developer says it invented. Read the teardown.

What they share

In several, the guard was inside the agent. In the git reset case the hook the agent said it made "was never actually created" (#34327). In the reboot case the developer says the context "got too long and it 'forgot' about all the rules and tools" (comment). In OpenClaw, approvals were a setting the stolen token could switch off; its docs say they "are not a per-user auth boundary" (OpenClaw docs).

Tego puts the limit where the agent cannot edit it: the Linux kernel and a supervisor outside the agent (README). A contract still refuses only what it names, so write lists short: the folders the task needs, the programs it runs and the hosts it reaches. Every allowed host stays a way out for data (Tego's known limits, item A1), and Tego sees nothing in systems that never run under it (Tego's known limits, item A5).

For the evidence behind the series, read AI agents now act on real machines. For every stated limit, read Tego's GAPS.md.

Try it: Tego on GitHub and Get started with Tego. Running agents across a team? Talk to us.

Sources

Sources checked on 5 November 2026.

Every link in this post, by site:

github.com/Attention-Span-AI/Tego
adnanthekhan.com
github.com/openclaw/openclaw
github.com/anthropics/claude-code
github.com/BerriAI/litellm
github.com/openai/codex
docs.openclaw.ai