Where GIRD stands on security and data

Updated 26 October 2026

This page sums up how GIRD handles data and security. The full detail is in our privacy notice and the Argus community edition privacy notice.

How GIRD decides

Argus runs on each Mac, as a background service and macOS network extensions. It watches your AI coding agents and decides what each one may do on the Mac itself: in its background service, in the macOS sandbox it runs agents under, and in its network filter.

Argus keeps a record only of activity that traces back to an AI coding agent it watches. Your other apps' traffic passes through only so Argus can trace it to its source, and is not stored.

Your files and your agents' conversations are analysed on your Mac. They leave it only if you turn on Ask Argus or a hosted scoring model. Scoring runs on your Mac by default.

Built to fail open

Argus is built so that a problem with Argus does not stop your engineers' work.

• If Argus is running but cannot protect an agent it watches for you, it starts the agent anyway, without protection, and logs that it did.
• If Argus is missing, for example in the middle of an update, the agent starts as normal.
• If the Argus background service is down, wrapped agents still start.
• The network filter blocks a connection only when your policy says to block it. An app it doesn't know, monitor mode or no policy at all lets the connection through.
• Inside Argus, each feature is built to fall back to earlier behaviour if it fails, rather than take down the service that watches your agents.

One exception, by design: when you start a protected launch yourself and Argus cannot set it up, that launch stops instead of running unprotected.

What GIRD collects, and what it never collects

GIRD collects data through this website, the Argus app and Blindspot, our AI key check script.

This website

• Our hosting provider records request logs, such as your IP address, your browser and the pages you request, to keep the site running and secure.
• The form you use to book a demo or a risk review sends your name, your work email, the AI agents you run if you add them, and the page you sent it from.
• The Blindspot update form sends your email address, used only to tell you when we add a new check.
• Our public pages use no analytics and set no cookies of our own. No advertising trackers, social-media pixels, session replay or cross-site identifiers.
• We never use your submissions to train any AI model. We do not sell or rent your data, or share it with marketers or data brokers.

The Argus app

Argus sends GIRD the data below so we can keep it stable, understand how it is used and improve its protection. This table describes Argus 1.3.30 and later.

Neither kind of telemetry ever includes a file path, a command, a prompt, a credential, or a host outside a public list of AI and developer services. The install ID is random, never derived from your hardware, name or email. While you're signed in, it is linked to your Argus account.

What the security log carries

Your Mac keeps the full log. Since Argus 1.3.30, the copy GIRD receives holds fixed fields only:

• Each event's kind, whether Argus allowed or blocked it, the agent and model, and the program's name. For a sign-in or a Keychain read, also its action, such as gh auth. The rest of the command is dropped.
• The destination: a well-known service by its name, such as api.anthropic.com, or by its domain, such as github.com. Any other site appears only as a keyed one-way hash of its domain. Private or intranet names and IP addresses are not sent at all.
• A file appears only as its category, such as a cloud credential, and a salted one-way hash of its path. Never the path itself.
• A secret Argus finds is named by its kind alone. So is an error, such as tls.
• Your Mac's hardware ID and computer name, only as salted one-way hashes, with the random install ID and the session and process IDs each event belongs to.
• Your email, while you're signed in.

Every field is checked against its type before it leaves, so a log line added later cannot carry arbitrary text either. One limit applies. The name of a program, model, MCP server or hook, or a word that says what happened, is sent as written when it reads as an ordinary word. One that looks like a token is sent as a one-way hash, or not at all. Even so, a short value an agent picks for one of these can reach the log.

Your files and your agents' conversations

The security log and telemetry never carry the contents of your files, or the text of what your agents send and receive. Two features you turn on yourself can send more, scrubbed first:

• Ask Argus sends your question and what Argus looks up to answer it. That can include a session's conversation, the commands it ran with their arguments, and the files and hosts it touched. Before anything leaves, Argus removes every secret, email address and IP address it recognises, and shows your home folder as ~. A secret it does not recognise can still be sent. GIRD's Ask Argus service passes the request to a model and keeps no copy of questions, answers or tool results. Your chats are saved on your Mac only.
• A hosted scoring model is off unless you set one up with your own OpenAI, Anthropic or Google key. That provider then receives the agent activity being scored, including prompts, commands, file paths and hostnames, scrubbed the same way, under its own terms.

Blindspot

When Blindspot, our AI key check script, finishes, it sends us one report, plus one small event if you use its menu.

• The report holds counts, such as coding agents, MCP servers and files with keys. It says which agents, key providers and well-known MCP servers it saw, from fixed lists, and where your keys sit by kind of file, such as shell history or .env files. It also holds a score from 0 to 100 with its grade, and how long the scan took.
• It holds a few facts about your setup, such as your operating system version, computer model, processor, terminal, shell and language setting.
• It includes a device ID: a one-way hash of your computer's hardware ID, salted for this script, so we can count repeat runs. It can't be turned back into the hardware ID, and we don't link it to your name or email.
• The menu event says whether you saved the report, opened your share card or opened the Argus download page. A share button on your result page tells us which button, what kind of result it shows and the run's random ID, nothing else.
• It never sends a file path, key, username, email, hostname or the raw hardware ID. We don't store your IP address, only the country Cloudflare derives from it.
• Run it with --offline, or set DO_NOT_TRACK=1, and it sends and asks for nothing.

We publish running totals across every device at gird.ai/api/check/stats, and never publish case-study figures about a group smaller than 20 devices.

Where data lives and how long it is kept

This is what our privacy notices state. Where they don't name a region, ask us at hello@gird.ai.

Argus rolled out by your organisation is covered by your organisation's signed agreement with GIRD.

To access, correct or delete your data, email hello@gird.ai. We respond within 5 business days, verify your identity before acting, and complete verified deletions within 30 days. For Argus, write from your account's email address. We then delete the data linked to that account, from every Mac you signed in on. Signing out stops new uploads but does not delete what was already sent. Under India's DPDP Act 2023, you can also raise a grievance at the same address, and we acknowledge it within 7 days.

Subprocessors

These companies handle data for GIRD.

Crash reports go to an error tracker GIRD runs on its own servers.

Services the Argus app contacts directly

• Public WHOIS servers get the registered domain of each public site your agents connect to, so Argus can flag newly registered domains. Private and intranet names are never looked up.
• Public threat lists, such as Spamhaus DROP, are downloaded to your Mac. Nothing about you is sent.
• Services you set up yourself, such as a hosted scoring model or a MaxMind geolocation download, are contacted only after you add your own key.

In the app, Settings › Privacy & Permissions lists every service Argus contacts, and why.

Security practices

Signed and notarized builds

Argus for Mac is signed with an Apple Developer ID and notarized by Apple, the app and the disk image both. The download page lists the disk image's SHA-256, so you can check your copy. Updates are signed too, and the app checks each one against a public key built into it.

Data goes only where it was built to go

A release build of Argus sends data only to the addresses it was built with. An environment variable that tries to send the security log, telemetry, account data, Ask Argus or crash reports anywhere else is ignored unless it names a GIRD domain over HTTPS.

Privacy rules held by tests

The fields the security log may send sit on a fixed allow-list. A change that adds a field fails the tests until the field is reviewed and added to the list, and our rule is that only hashed, coarsened or masked values get added. Each sentence the app shows about what leaves your Mac is listed with the rule that makes it true and the tests that check it.

A tamper-evident record on the Mac

Argus keeps its capture log, the record of what your agents did, as a hash chain. A deleted, reordered or edited entry breaks the chain and is flagged. This makes the log tamper-evident: someone with full administrator rights on the Mac could still rewrite the whole chain.

Access to Argus data at GIRD

The internal dashboards where our team reads Argus data sit behind Cloudflare Access, which lets in only the team email addresses on its list.

Responsible disclosure

Report a security issue to security@gird.ai. Under our disclosure policy:

• We acknowledge your report within 2 business days.
• We send a triage assessment, with a severity rating and a proposed timeline, within 5 business days.
• We fix high and critical issues within 30 days of triage wherever possible, and tell you if a fix needs longer.
• We update you when the fix is scheduled and when it ships, and credit you publicly if you want.
• If you act in good faith and follow the policy, we will not pursue legal action against you for your research.

The policy covers the Argus app in every edition, the Argus sign-in and telemetry services, the GIRD console, gird.ai and its subdomains, and anything else we distribute. We don't run a paid bug bounty. On request, we provide a written reference letter.

Security questionnaires

We complete your CAIQ, SIG or own questionnaire during onboarding.

Certifications

SOC 2: not yet. We share our current controls under NDA. We name a certification on this page only once we hold it.

Contracts and data processing terms

If you use Argus on your own, the Argus licence in our terms and our privacy notices apply. Team and Enterprise rollouts run under a separate signed agreement with GIRD, which prevails where it differs from our terms.

Ask us for our data processing terms at hello@gird.ai.

Contact

• Security reports: security@gird.ai. A PGP key is available on request.
• Everything else, including privacy requests and security questionnaires: hello@gird.ai.
• Our security.txt: gird.ai/.well-known/security.txt.