Anomity Now Detects Kiro Crew Bots - and Flags the Settings That Make Them Risky
- Anomity now detects Kiro Crew, the always-on gateway behind Kiro's crew bots: persistent bots that run on schedules and take commands from chat apps.
- The Endpoint Sensor reads Crew's own settings and flags the five things that turn a helpful bot into a standing risk: sandbox off, running unsandboxed allowed, every tool approval skipped, the dashboard published beyond the machine, and a chat app wired in as a command channel. A sixth flag catches bots that can use the person's SSH keys.
- Each one raises a finding under the Dangerous local agent runtime posture policy, with the reason written in plain language, and routes to your SIEM, Slack, email or Jira like any other finding.
- The inventory also notes whether a Crew has schedules configured and whether an administrator has set Crew's host governance policy on that machine.
- Crew completes our Kiro coverage: the Kiro IDE, the Kiro CLI and Kiro Crew are each listed under their own name, with their MCP servers, skills and the way each person signed in to Kiro.
- Metadata only. Anomity never opens Crew's
.env, signing keys, bot memory, sessions or work log, and a chat-app token is reported as a redacted finding, never as its value.
Most AI coding tools wait for a person. You open the IDE, you type a prompt, you approve a step. Kiro crew bots do not wait. Kiro Crew is an always-on gateway that keeps bots running on a laptop or a remote machine, wakes them on a schedule, and lets people talk to them from a chat app. That is exactly why teams want them, and exactly why a security team needs to know where they are.
Starting today, Anomity detects Kiro Crew on every managed device and flags the settings that make its bots risky. No new install and no configuration on your side: the Endpoint Sensor already on the fleet picks Crew up, and its findings land in the same console as everything else.
Why crew bots need their own coverage
We wrote about the governance shape of Crew when AWS first shipped it: the guardrails are good, and the autonomy model moves the problem. Three properties make crew bots different from the agents most security teams already track.
- Nobody is watching when they run. A scheduled bot acts at 03:00 on whatever it reads. Approval at execution is not in the path by design, so the settings configured beforehand are the whole control.
- A chat app becomes a command line. Crew can take instructions from Telegram, Discord, Teams, WhatsApp, iMessage and others. Whoever can message that bot can, in effect, run work on the machine it lives on.
- The dashboard can leave the laptop. Crew can publish its dashboard through a tunnel or a remote URL. Whoever holds a dashboard token can then drive the bots from the internet.
None of that is a flaw in Crew. It is how an always-on agent is supposed to work. But it means one person's configuration choices decide how much of their home directory, credentials and SSH keys an unattended bot can reach, and until now that choice was invisible outside the machine.
What Anomity flags
The Endpoint Sensor reads Crew's settings file and its local overrides, applying the same precedence Crew does, so a setting changed with a local override is judged by the value Crew actually uses. Each risky setting becomes a finding under the Dangerous local agent runtime posture policy:
| Finding | What it means |
|---|---|
| Sandbox off | Crew runs its bots with the OS sandbox off. Always-on agents get the whole home directory, credential folders included. |
| Running unsandboxed allowed | When a sandbox cannot be applied, Crew starts the bot anyway instead of refusing. |
| Every tool approval skipped | Bots act on schedules and chat commands with nobody approving a step. |
| Dashboard published | The tunnel is on, or the dashboard URL names a remote address, so the bots can be driven from beyond this machine. |
| Chat app command channel | A chat app is enabled as a way to send the bots commands: a remote command channel onto this machine. |
| SSH agent forwarding | Bots may use the person's SSH agent, and so every SSH key loaded in it. |
Each finding carries that reason in plain language, the device and the person, and follows the same routing as every other Anomity finding: SIEM, Slack, email or Jira, with the change kept in the 90-day audit trail. Crew also shows up under Inventory, so the first question, who is running crew bots at all, is a filter, not a survey.
Two more signals sit beside the findings. The inventory notes when a Crew has schedules configured, which is the line between a bot someone chats with and a bot that runs unattended. And it notes when Crew's host governance policy is present, which tells you an administrator has set limits on that machine rather than leaving every choice to the person running it.
One view of the whole Kiro suite
Crew is the third Kiro product Anomity covers, and each is listed under its own name, because each carries a different risk:
- Kiro IDE - detected as soon as it has been opened, with its MCP servers, project and user skills, and hook files. Kiro's own MCP config rewrite issue is why the MCP file matters.
- Kiro CLI - the terminal agent that succeeded the Amazon Q Developer CLI, told apart from the IDE even though they share one home folder, with its custom agents and the MCP servers they carry.
- Kiro Crew - the always-on gateway behind crew bots, with the posture findings above. The MCP servers its bots use come from the same Kiro settings file the IDE reads, so they are already in your MCP inventory.
Under AI Accounts, Kiro shows how each person signed in and their Kiro plan, so a personal plan on a company laptop stands out. Kiro records no email address, so Anomity reports the sign-in method and plan rather than guessing an identity. Put together, that answers the questions a CISO actually asks about a new AI suite: who has it, what is it connected to, and is any of it running unattended.
What Anomity does not read
Crew keeps sensitive material in its home folder, and the sensor is built to stay away from it. Never opened: Crew's .env, its token signing key, its trust keys, the bots' memory stores, sessions, the work log and skills. The schedules file and the governance policy are noted as present and never opened, because their contents are the person's own prompts and rules. The dashboard URL itself is never sent, only whether it points beyond the machine.
Crew can keep chat-app bot tokens inline in its settings. Like every other config the sensor reads, secrets are redacted on the endpoint: a token that matches a known pattern is reported as a finding with the pattern name and where it sits, never the value.
How Anomity fits next to Crew's own controls
Crew ships real guardrails: denied-by-default commands, sensitive-path blocking, OS-level sandboxing and signed audit logs. Anomity does not replace them. It answers the question those controls cannot answer from inside one product: across the whole fleet, which machines have turned them off? The same console covers the Claude Code, Cursor, Codex and OpenClaw installs on those laptops, so one policy for always-on agents covers every vendor, not just one.
Want to see which of your developers are already running crew bots, and how they are configured? Book a 30-minute demo and we will show you your own fleet.
Frequently asked questions
What are Kiro crew bots?
Crew bots are persistent AI bots run by Kiro Crew, Amazon's always-on Kiro gateway. They can run on schedules without a person present, and people can send them commands from chat apps such as Telegram, Discord, Teams or WhatsApp. Crew runs locally or on a remote machine and has a desktop app, a web dashboard and a terminal interface.
Does Anomity detect Kiro Crew?
Yes. The Anomity Endpoint Sensor detects Kiro Crew on managed devices and lists it in the inventory under its own name, separate from the Kiro IDE and the Kiro CLI. There is nothing to install beyond the sensor already on the fleet.
Which Kiro Crew settings does Anomity flag?
Six: the OS sandbox turned off, bots allowed to run unsandboxed when a sandbox cannot be applied, every tool approval skipped, the dashboard published beyond the machine through a tunnel or remote URL, a chat app enabled as a command channel, and bots allowed to use the person's SSH agent. Each raises a finding under the Dangerous local agent runtime posture policy.
Does Anomity read crew bot conversations or memory?
No. Anomity never opens Crew's .env file, signing keys, trust keys, memory stores, sessions, work log or skills. Schedules and the host governance policy are recorded as present, never opened. Chat-app tokens in the settings are redacted on the endpoint and reported only as a pattern name and location.
Does this replace Kiro Crew's own guardrails?
No. Crew's denied-by-default commands, sandboxing and signed audit logs work inside Crew. Anomity shows, across the whole fleet, where those controls have been turned off, and puts Crew next to every other AI agent and MCP server on the same devices so one policy covers all of them.
Which Kiro products does Anomity cover?
The Kiro IDE, the Kiro CLI and Kiro Crew, each under its own name, with their MCP servers, skills, hook files and agent files, plus how each person signed in to Kiro and their Kiro plan under AI Accounts.




