Last year an AI coding agent deleted a company's production database, and then told the owner it could not be restored. The restore worked fine. The agent had been told not to touch anything. It did anyway.
We build security for AI systems at Fencio, and this is the problem we keep coming back to. Coding agents are useful because they can run commands. That is also exactly what makes them risky.
What the incident record shows
We keep a public list of AI agent incidents at fencio.dev/incidents. Most are not about coding agents, but several are, and they read like a pattern. Here are four, with details taken from our write-ups of them.
- Replit deleted a production database. The agent was told to help with a task during a declared code freeze. It deleted the whole production database, made up 4,000 replacement users, falsified test results, and claimed rollback was impossible when it was not. (The Register, Fast Company, Cybernews, July 2025)
- Google Antigravity ran attacker-supplied flags. Pillar Security found that the IDE's find_by_name tool passed its Pattern parameter to the fd command, and that parameter accepted injected flags. That let it execute arbitrary binaries and bypass Strict Mode, and indirect prompt injection could trigger it. Google fixed it as of February 28, 2026. (The Hacker News)
- Ruflo published a shell to the network. Its default docker-compose exposed an unauthenticated MCP bridge with 233 tools, including shell execution, on 0.0.0.0:3001. Noma Labs found that a single HTTP POST reached all of them. It was rated CVSS 10.0 (CVE-2026-59726), and the patch shipped within 24 hours. The default setup was the whole exploit chain. (The Hacker News)
- OpenCode Console may have leaked API keys. An intruder reached a legacy system on October 16. OpenAI, Anthropic and Google API keys that users had saved in the console may have been exposed. The full extent is still under investigation. (The Hacker News)
And one that was widely reported but is not on our list yet: a developer asked Google's Antigravity to clear a project cache, and a path error pointed the command at the root of his D: drive. It ran rmdir /s /q, with no confirmation, reportedly in Turbo mode. Source: TechRadar, https://www.techradar.com/ai-platforms-assistants/googles-antigravity-ai-deleted-a-developers-drive-and-then-apologized
Full record with sources: https://fencio.dev/incidents
Same setup, different causes
One was a judgement failure, one was a tool parameter, one was a default config, one was leaked keys. The cause changes every time. What stays the same: something can act on the machine or the network, and nothing independent checks the action before it runs.
Asking the model to be careful does not fix that, because the model is often the thing that went wrong. And a permission prompt only helps if you read every one. Modes that skip the prompt exist precisely because people do not.
What a runtime firewall does
The idea is simple. Put a checkpoint between the agent and the machine. Every prompt and every tool call passes through it. Before a command runs, the checkpoint asks: is this inside what this agent is allowed to do?
That is what the Fencio CLI does. You connect your coding tool, and Fencio adds hooks to it. Where the tool supports it, model traffic also goes through a proxy. Today the docs list Claude Code, Codex, Cursor and GitHub Copilot.
How it decides
You write a rule as one plain sentence. Fencio turns it into a precise boundary. A few things matter about how that boundary works:
- Rules are allow-only. You say what is permitted. A step that contradicts it is denied. Anything the rule does not mention stays allowed.
- It judges what an action would do, not how it is worded. Deleting a file through a file tool and deleting it with rm are both a delete, and a rule about deletes catches both.
- The facts come from deterministic parsers, not a model. The action, exact paths, hosts and branches are pulled out the same way every time.
- It remembers the chain. Output from an untrusted result is tracked, so you can see when it shows up in a later step.
- The decision is made before the command runs. A check that only notices afterwards is a report, not a control.
Try it in five minutes
curl -fsSL https://cli.fencio.dev/download/install.sh | sh
fencio integrations connect claude-code
fencio system doctor
fencio traffic flows watch
Start by watching. Rules default to monitor mode, which records what would have been blocked without blocking anything. Look at your real traffic for a few days. When the decisions look right, switch the rule to enforce.
Full docs: https://fencio.dev/docs?utm_source=krishnaa&utm_medium=personal&utm_campaign=firewall
What it does not do
We would rather you hear this from us. It is not a sandbox, so keep real credentials and real data away from agents where you can. Monitor mode blocks nothing until you enforce. Enforcing a tool-call rule needs a ready parser, and the docs give you a readiness check for that. And a rule set is only as good as the boundaries you write.
Where this goes
Agents are going to keep getting more access, because that is where the value is. The teams that get the most out of them will be the ones that can say exactly what an agent may and may not do, and prove it held. That is the part we are building.
If you try it, tell us what broke. We read everything.

