Arcezia / AI agent security
AI agent security: controlling what agents do
AI agent security covers three different jobs. Protecting the model: stopping prompt injection and jailbreaks from changing what it says. Protecting data: stopping leaks and keeping private information private. Controlling what an agent actually does: the tool calls it makes, the approvals those calls need, and actions that cannot be undone. This page is about the third. One SQL call, checked three times on the live service, shows what that control answers: allowed, held and refused.
Why the third job is different
A correct-sounding answer and a harmful tool call can come in the same turn. The first two jobs mostly look at content: what goes into the model and what comes out. The damage is often in the call: which record, which account, how much, which environment.
In the Replit incident, the agent acknowledged a code freeze and then ran destructive commands against the live production database. In the Lobstar Wilde incident, the agent decided to send a small amount and the call it made sent a thousand times more.
In the agent-incident-ledger, 5 of 6 incidents were tool calls: the call itself did the damage. In two of them the agent had been told plainly not to do it. The count is small and it is not a sample of anything. All six incidents.
What control at the call looks like
A check sits between the agent deciding to call a tool and the tool running. It answers one of three ways:
- ALLOW: the call runs.
- REVIEW: the call is held, and the answer names what would release it, such as a person’s approval.
- BLOCK: the call is refused, and the answer says why.
A person’s approval is bound to one call and used once, so it cannot be stretched to cover the next call. An approval the agent only claims does not count. Each decision is kept as a signed record with its reason.
See it on one call: the same SQL read allowed, held and refused, with the code and the answers measured on the live service. Check the records yourself: a signed sample export and the offline checker.
Agent runtime security
Agent runtime security is control of what an AI agent does while it runs: each tool call is checked at the moment the agent makes it, before the tool acts, rather than found in logs afterwards. It is different from testing an agent before release, and from keeping an inventory of agents and their settings. At runtime the question is narrow: may this call, with these arguments, run now? The same job is also called execution security, action authorization or pre-execution control.
Arcezia answers that question with ALLOW, REVIEW or BLOCK, and the call reaches the tool on ALLOW alone. It works where the agent runs: a hook in Claude Code that also covers MCP tool calls, and the framework integrations. One call checked three ways on the live service: action authorization.
AI security for agents
AI security for agents is AI security for chat models plus one more question, because an agent acts. Defences against prompt injection, output filters and data controls still apply to what the model reads and writes. What they cannot settle is whether a tool call that reads as harmless should run, since the harm is in the arguments and the effect: which table, which recipient, how much.
So a stack for agents keeps its guardrails for text and adds a check at the tool call, with rules you write and a signed record of each decision: agent governance, end to end. How the layers fit together: where Arcezia fits.
Definitions
- A tool call
- A tool call is the request an AI agent makes to a real system: run a command, delete a record, send an email, move money.
- Controlling what an agent does
- Controlling what an AI agent does means deciding whether each tool call may run, before it runs.
- Pre-execution verification
- Pre-execution verification is checking a tool call an AI agent proposes, and the facts it rests on, before the tool runs, then answering ALLOW, BLOCK or REVIEW.
- An approval bound to one call
- An approval bound to one call releases that call once. A different call, or the same call again, needs a new approval.
- A signed decision record
- A signed decision record is the receipt kept for each check: the answer and its reason, signed so that a later change to the record shows when it is checked.
More terms: glossary.
If you build agents
How do I stop my agent running a destructive command?
Put a check between the agent and the tool, so the command is answered before it runs. For Claude Code, after pip install -U arcezia and setting your key, arcezia-hook install adds a hook to ~/.claude/settings.json that checks each shell command, file write, web request and MCP call first. Then add "onFailure": "block" to that hook entry, so a missing or crashed hook blocks calls instead of letting them through (Claude Code 2.1.295 or later). Claude Code setup.
LangChain and LangGraph, n8n and the other frameworks have their own setup pages. On a new key a call usually comes back REVIEW first, and the answer names what would release it.
If you run security
What stops an agent with valid credentials from doing damage?
Credentials answer who may act. They do not answer whether this call, with these arguments, should run now. In the PocketOS incident the token had been created for adding and removing custom domains but was scoped to any operation, and one delete call spent it.
An outside architect tested the public checking service. The report is published in full, including what it says it does not establish: “production security, regulatory compliance, certification, or suitability for deployment.”
If you answer to auditors or regulators
Can I show who approved what, and why each action was allowed?
Each decision gets a signed record with its reason, made before the action runs; a later change to a stored record shows when it is checked. A person’s approval is a token your backend signs after the person clicks Approve, bound to one call and used once. An approval the agent merely claims is refused.
The developer docs map these records to EU AI Act Articles 9 and 11 to 15: Receipts and compliance. The record is evidence, not a certification, and using it does not by itself make a system compliant.
If you run the business
What could my agent cost me?
The incident pages record what was lost and what came back. In 5 of the 6, the damage was done by a tool call, not by what the model wrote.
- OpenAI agents and Hugging Face: code ran on 41 production workers; Hugging Face rotated every token, credential and signing key and rebuilt core infrastructure.
- PocketOS production volume: three months of bookings and payment records gone; restored first from a three-month-old backup, then by Railway.
- OpenClaw inbox deletion: hundreds of emails trashed or archived; the agent stopped only when the processes on the host were killed.
- Lobstar Wilde token transfer: tokens sent to a stranger's wallet that cannot be recalled; no recovery has been reported.
- Replit production database: production data deleted; it came back through Replit's own rollback, the one the agent had said was impossible.
- Air Canada chatbot: a tribunal award of 812.02 Canadian dollars against the airline; no tool was called, the harm was in the words.
If you research or report on this
Where is the data?
Every incident comes from the agent-incident-ledger on GitHub, CC-BY 4.0, one file per incident. A file needs at least one primary source: the operator’s own post, the vendor’s statement, a court or regulator document. Press coverage is listed as secondary, never alone.
To dispute a classification, open an issue on the repository. To add an incident, open a pull request.
If you use AI agents
Is my agent safe to let act for me?
Ask three questions of any agent tool before you let it act for you.
- Can it be stopped before an action that cannot be undone?
- Does a person approve the risky steps?
- Is there a record of what it did and why?
In one incident on this site, stop messages sent from a phone did not halt the agent; only killing the processes on the host did.
How Arcezia does this
Arcezia puts this check in front of an agent’s tool calls: ALLOW, REVIEW or BLOCK before the tool runs, approvals bound to one call, and a signed record for each decision. Developer docs.
Check it yourself
- Action authorization: one tool call, three answers
- Audit evidence: signed records you can check offline
- Where Arcezia fits next to guardrails, observability, evaluations, orchestration, identity and sandboxes
- AI guardrails: what text checks cannot see, measured
- MCP security: MCP tool calls checked in Claude Code
- Agent governance: a rule, a decision, a signed record, an offline check
- Records for banks and for healthcare, with audit procedures
Incidents
- 8 July 2026: OpenAI evaluation agents broke into Hugging Face production
- 24 April 2026: Cursor agent deleted PocketOS's production volume on Railway
- 23 February 2026: OpenClaw agent trashed an inbox it was told not to act on
- 22 February 2026: Lobstar Wilde agent sent a thousand times more than intended
- 18 July 2025: Replit agent deleted a production database in a code freeze
- 14 February 2024: Air Canada chatbot misstated its bereavement fare policy
All incidents, with how they are classified.
Integrations
- Claude Code
- LangChain and LangGraph
- OpenAI function calling
- Anthropic tool use
- n8n
- CrewAI, AutoGen and LlamaIndex