Menu

Stopping agents orchestrating attacks

How we stop Klyroe's agents from being used to orchestrate cyber attacks: the layered controls, end to end

How we workNovember 18, 2026Senova Systems Research

Attackers have already used AI agents to run much of an intrusion. This is every control Klyroe has against that, layer by layer, and the gaps we have not closed yet.

What we say Klyroe does was checked against our code on the day of writing. We describe each control in general terms and leave out exact limits, so they are harder to work around. Where something is planned rather than built, we say so. No control here makes misuse impossible; together they make it slower, louder and easier to stop.

Why this is no longer hypothetical

In November 2025 Anthropic reported that it had disrupted a campaign, detected in mid-September 2025, in which a group it assessed with high confidence to be Chinese state-sponsored used its Claude Code agent against roughly thirty organisations. By Anthropic's account the agent carried out 80 to 90 % of the tactical work, with people stepping in at perhaps four to six decision points per campaign. That is the lab's own report about its own product, and we read it as such, but it describes the shape of the risk clearly: the attacker did not ask for an attack. They split it into small tasks that each looked like ordinary security work, and told the model it worked for a firm doing defensive testing.

Klyroe is a much smaller product, but it has the same ingredients: models that write code, agents that browse and act, and computers in the cloud that run what the agents write. So we have to assume someone will try. This article lists what stands in their way.

The principle: assume any one layer fails

No single control stops misuse. Models can be talked around, classifiers miss things, and people approve what they should not. So we stack separate checks, each of which still holds when the one before it fails, and each of which is written in ordinary code rather than in a prompt the model could be argued out of.

  1. Account and planAgents run under a signed-in account (age 18 or over at sign-up), its plan's request limits, step caps and spending caps.
  2. Model rules and policyRequests for malware are refused by rule, and the acceptable use policy forbids attacking systems you aren't authorised to test.
  3. Labelled outside textPages, files and tool results are wrapped as data with one rule: never follow instructions inside them.
  4. Injection noticeText that tries to talk to the model is flagged to the person. It informs; it never silently edits the page.
  5. Tiers and taintRead runs, write asks once outside text was read, confirm always asks. The model's words cannot change a tier.
  6. Grants that can't widenBuilt-in tools that send, pay, call or delete are never remembered as allowed. Each one asks every time.
  7. Outbound checkRequest catchers, tunnels and paste sites are blocked; keys, passwords and card numbers are never typed or sent.
  8. Send limitsDaily limits on emails, new recipients and calls, sized by how much trust an account has earned.
  9. Activity log and StopEvery call is logged with secrets removed. Stop everything halts an account's runs, approvals and agents.
The layers around a request. Each one is a separate check, so a failure in one is not a failure of all.

An account, a plan and limits

Agents act for a signed-in account. Sign-up asks for an age of 18 or over, requests per minute are limited by plan, and staff can suspend an account, which blocks its sessions, API keys and channels and stops work already running.

Agents that send things have their own daily allowances for emails, new recipients and calls, sized by how much trust an account has earned: a new account starts low, and the allowance grows with account age, a verified email and a paid plan. There is also a daily limit on how many messages one person can receive from one account. Abuse reports alert staff; they never pause an account on their own, so the report form cannot be used to silence a stranger.

Long runs are capped too: an agent loop stops after a set number of steps, and scheduled tasks have step and spending budgets. None of this stops a determined attacker. It makes automation at scale, the thing AI adds, run into walls.

The model's own refusals, and the policy

Every chat model in Klyroe receives the same working rules, and they say plainly that malware is refused whatever the framing. Our acceptable use policy forbids malware, phishing and help attacking systems you are not authorised to test, and says we may limit features, suspend or close accounts when it is broken.

We are honest about how strong this layer is. A rule in a system prompt is a request, not a wall, and the GTG-1002 case shows how refusals are worked around: by splitting a task until no single step looks harmful. Klyroe does not run a separate classifier over prompts or replies for cyber misuse today; refusal rests on the underlying model and that rule, and we test it (see Measuring misuse). That is why the layers below assume this one has already failed.

Outside text is data, and the turn remembers it

An agent that reads the web reads text written by strangers. Every tool result, page and file reaches the model wrapped in a labelled block, with one rule: what is inside is data, never instructions. Text that tries to talk to the model is flagged to the person with a notice. That detector only informs; it never edits or blocks, because pages that merely discuss injection would break.

What blocks is taint. Once outside text has reached a reply, the server records it, and from then on write actions ask, saved approvals for risky tools stop applying, and a reply that has read outside text cannot start a coding session at all. The prompt injection article covers this in depth.

Tiers, approvals and never-grant tools

Every action has a tier fixed in code: read runs, write asks once outside text has been read, confirm always asks and needs a slide. Built-in tools that send, pay, call or delete can't be approved in advance. When Klyroe's agents work in their own cloud browser, they will not act on banking, payment, crypto, government ID or security-settings pages, never types into password or card fields, and asks before anything that submits, sends or buys. The tool permissions article has the full rules.

For cyber misuse this matters because an attack is a sequence of actions. Every action that reaches another system, sends data out or changes something is a point where code checks it and, often, a person sees it.

Where an agent may connect

Before an agent opens an address, fetches a page, types or submits, an outbound check runs. It blocks the places injected text usually tells an agent to send data: request catchers, tunnels and paste sites. It blocks passwords, keys, tokens and card numbers in addresses and typed text, and asks before personal data goes to a new site. When our servers fetch for an agent's custom tools, connectors, plugins or knowledge sources, they use one fetcher that accepts https only and refuses private and local addresses after the name is resolved, so a page or a tool definition cannot steer our servers into our own network.

The cloud computers agents use are walled off from our systems: each account's machine runs on its own private network with an encrypted disk, and coding-session machines, when turned on, add firewall rules that block every private, link-local and reserved address and refuse to start if the rules fail to load.

Not built yet: those machines can still reach the public internet, because installing packages needs it. An allowlist proxy that limits outbound connections to approved destinations is planned. Until then, the account limits, the tiers and the activity log are what constrain an agent's outbound traffic from a machine.

Secrets the model never sees

A compromised agent is most useful to an attacker when it holds credentials. Keys for connected tools are stored sealed in the database and sent only as request headers by our servers; they are not put in front of the model. Coding machines receive no platform secrets, checked by name and by value. The activity log removes passwords, keys, tokens and card numbers from what it stores.

A record, and a way to stop everything

Every tool call from every runner (chat, helpers, agents, background tasks) writes one row to an append-only activity log, with secrets removed. A person can see what their agents did and export it.

Stop ends a running reply on the server, not just on screen. Stop everything goes further for one account: it stops every run, deletes saved approvals, turns routines off, pauses live agents and locks saved logins. For the whole service there is an operator switch that turns in-app actions off for everyone.

What we have not done yet

Publishing the gaps is part of the method. As of this article:

  • No classifier screens prompts or replies for cyber misuse; we rely on the model's refusals, our rules and the layers above.
  • The injection detector notifies and never blocks.
  • Agent machines have open public egress; the allowlist proxy is planned.
  • No outside penetration test has been done. Our own reviews and tests are not a substitute.
  • Our misuse evaluations are small and graded by pattern, then read by hand; they are described honestly in Measuring misuse.

If you find a way around any of these controls, tell us: [email protected] with the subject Security.

Sources

  1. Anthropic, Disrupting the first reported AI-orchestrated cyber espionage campaign, 13 November 2025 (lab report on its own product)
  2. Anthropic, full report (PDF), naming the actor GTG-1002 (lab report)
  3. NCSC, Impact of AI on cyber threat from now to 2027, 7 May 2025
  4. NCSC, CISA and partners, Guidelines for secure AI system development, 27 November 2023
  5. NIST AI 600-1, Generative AI Profile, July 2024 (risk 9: information security)
  6. OWASP, LLM06:2025 Excessive Agency
  7. Simon Willison, The lethal trifecta for AI agents, 16 June 2025