What Pigeon collects, where it goes, and how long it stays.
Effective 2026-10-10 · Operated by [OPERATOR NAME] · Contact privacy@pigeonlabs.academy
- Pigeon is an invite-only beta for education and research. Nothing is sold, and nothing is used for advertising.
- The sensor collects security telemetry from machines you install it on. Your account's data lives in its own database and event lake.
- AI prompts are sent as a fingerprint and a length, not as text, unless you switch capture on.
- Event history is kept for 365 days, then deleted.
1. Account information
When you are invited, we store your email address, name, role, and the hashes of your password and invitation token. Your second factor is stored as a passkey public key or an authenticator secret. Sign-in sessions are kept so you stay signed in. Accounts exist only by invitation.
If you request an invite on the Beta page, we keep your name, email address, what you do and what you wrote, to decide on the request, plus a one-way fingerprint of your network address for an hourly limit on requests. You get a short acknowledgement by email, and the operator gets your answers by email.
If you write to a Pigeon address (at pigeonlabs.academy or pigeon.observer) or use the Contact page, your message (including any attachments) is received through Resend and kept as a case on Pigeon’s support desk, where the operator and the people helping them can read and answer it. Replies and notes are kept with it until the case is deleted.
2. What the sensor collects
On machines you choose to protect, the sensor collects:
- Activity: program starts and exits, file creation and renames, network connections, DNS on Windows, persistence changes, and security verdicts from the operating system.
- Inventory: installed software and OS state, local accounts and logons, browser extensions, AI apps, MCP servers, and model files on disk. Model files are fingerprinted, never uploaded.
- Machine facts: hostname, OS and hardware model, network interfaces and addresses, MAC address, network neighbours, and the sensor's own health every 60 seconds.
- AI agent activity: the tools an AI agent calls, joined to the events above. With the optional browser extension, which AI websites are used.
What it does not collect: keystrokes, screenshots, clipboard contents or environment-variable values, which are recorded by name only. AI prompt text is reduced to a fingerprint and a character count on the machine. Text is sent only if you switch prompt capture on for that machine, or open a time-limited investigation window, which strips secrets first. File contents leave a machine only when an administrator asks for a specific file through the response terminal. Credential stores and key folders are always refused. Command lines are scrubbed of secrets on the machine.
3. How it is used
To detect and explain security events on your machines, send the alerts and reports you ask for, answer your questions through the assistant, and keep the service running. Data is not sold, shared for marketing, or used to train AI models by us.
4. Services that process data
- Cloudflare: receives sensor data and stores databases and the event lake, in North America.
- Vercel: hosts the console, in the United States. Its AI Gateway relays AI requests.
- Anthropic, Z.ai and OpenAI, through that gateway: alert explanations and the mail judge receive alert details. The assistant receives your questions and the data it reads to answer them. Voice features process your audio. The alert explanation step cannot be switched off.
- Resend: sends email from send@pigeonlabs.academy.
- Pushover: push notifications, where enabled.
- GitHub: hosts sensor downloads.
The site and console use no advertising or analytics trackers.
5. How long it is kept
- Event lake: 365 days, rolling.
- AI activity and health samples: 365 days by default. An administrator can shorten this to as little as 30 days.
- Live event view: 30 minutes.
- Captured prompt text: 7 days by default, at most 30. It is never written to the event lake.
- Alerts, hosts, settings and accounts: until they are deleted, or the account is closed.
6. Separation and security
Each account has its own database and its own event lake table. A shared control plane holds only account records, membership and software release catalogs. Every user has a passkey or an authenticator app. Response commands are signed by the console and verified on each machine. As the operator of a hosted service, [OPERATOR NAME] can technically access stored data, and does so only to run, secure or repair the service, or when you ask.
7. Your choices
You decide which machines run the sensor, whether prompt capture or prevention is on, and which emails you receive. You can remove the sensor with the uninstaller at any time. To see, export or delete your data, or close your account, write to privacy@pigeonlabs.academy.
8. Other people's data
Install Pigeon only on machines you own or have permission to monitor, and tell the people who use them. You are responsible for that permission.
9. Changes
This notice changes when Pigeon changes. Material changes are announced by email to account holders before they take effect.
The browser extension
Hawkeye Browser Sensor is part of Pigeon, an endpoint security agent operated by the organisation that manages your computer. It runs only on the AI websites listed in its manifest.
What it reads
For each prompt you send to one of those sites, it looks for credentials, source code or personal data. Everything is processed on your computer.
What it reports
For each prompt, it reports the category found, a confidence score, the site, and a one-way SHA-256 hash of the prompt. The text of your prompts is never sent anywhere. The report goes to the Pigeon sensor installed on the same computer, which forwards it to your organisation’s Pigeon console as security telemetry.
What it can change
If your administrator enables redaction, credentials such as API keys and passwords are removed from a prompt before it is sent. If blocking is enabled, a prompt carrying a credential is stopped, and you are told why.
What it does not do
Ad watch (only if you allow it)
If your organisation (in a family, your parent) turns on ad watch, the extension’s popup asks you whether to allow it, and nothing changes unless you click Allow. With it allowed, the extension watches the frames ads load in, on every site, for two things ad-borne spyware does: a frame bouncing through many sites, and a frame opening a page by itself. For those it reports how many sites the frame passed through and one-way SHA-256 hashes of the sites’ names (for example of example.com, never example.com/your/page). It never reports an address, a path or a query, or anything about the page you are on.
Apart from ad watch, it does not collect browsing history or read other websites. It never sells or shares data, shows advertising, or sends anything to its developer. It has no network destination other than the sensor on your own computer.
Questions about how your organisation uses this data go to your IT or security team. Questions about the software go to privacy@pigeonlabs.academy.
What leaves your machine: the whole of Pigeon
The extension is one part. This is what the Pigeon sensor and console send, keep and share, stated from the code.
What the sensor sends
The sensor sends to your own console only: the ingest worker, command and content endpoints, and pinned downloads (sensor updates, osquery, Sysmon). Nothing goes to the developer.
- Programs started. Path, command line and arguments, user, parent process and its command line, and (macOS) the code signature. From sensor 0.4.63 every command line is scrubbed on the machine before it is sent: API keys, tokens, passwords and private keys typed on a command line are replaced by [REDACTED:<kind>] and the event says how many were removed. Local prevention rules still read the original on the machine; nothing else does. Sensors before 0.4.63 send command lines as they are: a secret typed on a command line there reaches the console.
- Network connections. Protocol, local and remote address and port, the remote domain (looked up by reverse DNS unless reverse_dns is off), bytes in and out.
- Files. Path, kind of access, extension and size. Never contents, with one exception: someone with response access can read a file with RTR cat (at most 64 KB; keychains, ~/.ssh, ~/.gnupg, ~/.aws/credentials, private keys and the sensor's own folder are always refused), and RTR ls, ps, runscript and osquery send their output (at most rtr_max_output_kb, 1 MB by default). Every RTR command is recorded.
- AI agent prompts. A SHA-256 hash (first 128 bits) and a character count for each prompt, response, tool result and thinking block. The text itself in two cases only, both with API keys, tokens, passwords, private keys and connection-string passwords removed on the machine first: inside an investigation-capture window that someone with policy.manage opened for that machine or its group, or that an alert opened where a group turned auto-arm on (prompts, responses, thinking and tool results, at most 16,000 characters each; every window ends by itself within 72 hours and records who opened it and why); or when capture_prompt_text is turned on in that machine's own config.json (off by default, at most 8,000 characters, never thinking). capture_deny in the machine's own config refuses every capture window.
- AI tool calls. The tool name, a hash of its arguments, and one short hint (the command, a path or a URL) with API keys, tokens and passwords scrubbed, at most 300 characters. Also the session title, working folder and model name.
- Prompt-injection verdicts. The rule, a confidence, hashes of the matching span and of the body, and a character count. Never the text.
- PowerShell script blocks (Windows). Indicator names, a hash, the length and the script path. Never the script text; a block with no indicator sends nothing.
- Browser extension. Categories found (credentials, code, personal data), counts and hashes, to the sensor on the same computer only. Never the prompt text, and it has no other network destination.
- Environment variables and MCP servers. Environment variable and header NAMES, never their values. An MCP server's command, arguments and URL are sent as written in its config file, with API keys, tokens and passwords removed on the machine first (from sensor 0.4.63).
- Inventory. Installed applications and software, browser extensions, AI model files (path, size, the file's own header, a hash of three 1 MiB samples; never the weights), local accounts and logons, devices on the local network, privacy permissions (macOS TCC) and persistence items.
- Heartbeat (every 60 seconds). Sensor version, OS and hardware model, MAC address, local IP and network interfaces, and the sensor's own resource use.
- Never collected. Screenshots, the clipboard, keystrokes, the values of environment variables, and file contents outside an RTR request.
Where it is kept, and for how long
- The lake (Cloudflare R2). Every event the sensor sends, in full except captured text (the ingest worker removes it from every event before the lake), in the R2 bucket hawkeye-telemetry (Iceberg table hawkeye.events, written by Cloudflare Pipelines). Retention: 365 days, rolling. A daily job deletes events ingested more than 365 days ago (whole days, through the table), and the files left behind are removed after 3 days; the first deletions happen a year after the lake began (2026-09-06).
- The control plane (Cloudflare D1). The database hawkeye: machines, alerts, inventories, AI activity and sessions, commands, health samples, mail queue and Dot chats. Alerts are deleted 90 days after they were detected, whatever their status, once their alert.created record (and, if they were decided, their alert.closed record: final status, who and when, the decision log) is in the lake, which keeps those records 365 days. AI activity, health samples and hourly resource use keep 365 days. Other rows go when you delete a machine permanently, delete a Dot chat or unsubscribe.
- Captured text (Cloudflare D1, captured_text). Text from investigation-capture windows and capture_prompt_text, already scrubbed on the machine. Never in the lake. Deleted by a job every 15 minutes once it is older than the retention set on Policies › Machine groups: 7 days by default, at most 30.
- Short-lived stores. Command queue in Cloudflare KV (12 hours; download chunks 1 hour). Per-machine duplicate detection kept 7 days, ingest counters about 400 days, and the live stream (each machine's last 30 minutes of events) in its Durable Object.
Which AI processes what
Every AI call goes through the Vercel AI Gateway, which routes it to the provider of the model named: Anthropic, OpenAI or Z.ai. The worker on the ingest path calls no AI.
- Alert judge (a second opinion on every alert scoring 60 or more, and on sensor schema drift). anthropic/claude-sonnet-5. Gets: The alert: rule, host name, AI agent, summary and evidence, plus up to 40 recent events from that machine (program names, command lines cut to 200 characters, file paths, domains and IP addresses, tool hints). Turn off: Cannot be turned off from the console yet. Alerts below 60 never reach it.
- Mail judge (send or hold an alert email), its notes on repeated noise, and the daily review memo. zai/glm-5.3-flash. Gets: The alert's rule, host name, agent, summary and command line, and the household's recent mail history. Turn off: Turning off the review memo and the noise notes (Notifications) stops those two calls; the send-or-hold call runs for every alert email.
- osquery: plain-language question to SQL. zai/glm-5.3-flash. Gets: Only what you type (at most 500 characters) and the OS list. Turn off: Runs only when you ask a question in plain language.
- Dot, the assistant (chat, memory summaries). openai/gpt-6-luna. Gets: Your name, email address and role, the page you are on, Dot's memory of earlier chats, and the results of what she looks up for you (which can include alert evidence, command lines and lake rows). Turn off: Remove the assistant.use permission from a role (Users & roles).
- Dot by voice. openai/gpt-4o-transcribe, openai/gpt-4o-mini-transcribe, openai/gpt-4o-mini-tts, openai/tts-1-hd, openai/gpt-realtime-2.1. Gets: Your voice, her reply text, and the same context as chat. Turn off: Used only when you talk to her; the same permission turns it off.
- Adversary Intelligence reports (Premium). zai/glm-5.3. Gets: Public intelligence only (MITRE ATT&CK, MISP, news, vendor reports). No telemetry from your machines. Turn off: Runs only for reports you subscribe to.
- Automation watchdog: a three-line diagnosis of a failed internal pipeline from its log; receives no customer data. anthropic/claude-haiku-5.5. Gets: The failed check's own finding and, for a failed GitHub Actions run, the last 150 lines of that job's log. Nothing from your machines or your account. Turn off: Runs only when one of Pigeon's own automations turns red, for Pigeon staff.
- Exploited on your fleet: two or three plain sentences saying what a known-exploited CVE is, for the alert email. anthropic/claude-haiku-5.5. Gets: Only the CVE id and its public NVD description. Never a machine name, an address or anything else from your account. Turn off: Notifications: switch off Known-exploited CVE on your machines; without the model the email uses the NVD description.
- Threat-intel pipeline (news reading and story grouping, reports, rule drafts and their check; runs in GitHub Actions). anthropic/claude-sonnet-5, anthropic/claude-haiku-5.5, zai/glm-5.3-flash, openai/gpt-5-nano. Gets: Public feeds only. No telemetry from your machines. Turn off: Skipped when no gateway key is set for the pipeline.
Email, push and other services
- Resend. Alert emails (rule, host name, agent, summary, command line cut to 160 characters, the judge's verdict), the daily digest, health and review emails, known-exploited CVE alerts (the CVE, the machine names and installed versions, the fix and a link to the drafted patch job), intel briefs you opted into: to the household's recipients. Turn off: Notifications: switch off every kind except critical alerts, which always go.
- Pushover. A push per notified alert (severity, score, alert name, host name, agent, summary) and the daily digest's counts, when the deployment sets a Pushover key. Turn off: Only by removing the deployment's PUSHOVER_TOKEN; there is no console switch yet.
- Cloudflare. Everything the sensor sends: the ingest worker, the D1 database and the R2 lake are Cloudflare services (account region ENAM). Turn off: This is where the product runs; it cannot be turned off.
- Vercel. Hosts the console (region iad1) and the AI Gateway. The console has no analytics, tracking or error-reporting SDK. Turn off: This is where the console runs.
- Exa, Tavily and Perplexity (search). The threat-intel pipeline's daily news search: adversary names and the words of the query only, from GitHub Actions. Nothing from your machines or your account. Turn off: Only by removing the pipeline's EXA_API_KEY, TAVILY_API_KEY and PERPLEXITY_API_KEY; Pigeon then searches Google News.
- GitHub. Machines download sensor updates and osquery from GitHub releases (pinned SHA-256); the console reads release notes. No telemetry is sent. Turn off: Sensor updates: Sensor updates policy.
- VirusTotal. Only if your admin adds your own VirusTotal API key (Administration › Integrations): the file hash (MD5, SHA-1 or SHA-256) of a file in an alert, detection or quarantine, under your key, when someone opens it or clicks Check VirusTotal. Never the file, its path or the machine's name. Each answer is kept 24 hours in your own database. Turn off: Administration › Integrations: remove the key (the VirusTotal and MalwareBazaar links stay; they open only when you click them).
How to turn things off
- On the machine (its config.json; the console cannot change these). capture_prompt_text (off by default) keeps prompt text off; capture_deny refuses investigation-capture windows; prompt_flow false stops prompt monitoring entirely; reverse_dns false stops domain lookups; watch_sensitive_files (off by default) controls sensitive-file reads; rtr_deny_paths adds folders RTR cat refuses.
- In the console. Investigation capture (Policies › Machine groups, and each machine's Sensor health page): off by default; open or end a window, turn auto-arm on or off per group, and set how long captured text is kept. Notifications: which emails go (all but critical alerts can be switched off). Browser AI policies: observe, redact or block. Intel reports: the weekly brief and premium reports are opt-in. Users & roles: remove assistant.use to turn Dot off for a role.
- Not possible yet. The alert judge cannot be switched off, and Pushover has no console switch. These are listed so the page does not promise them.