ChatGPT agent and Atlas in server logs: how to spot them
ChatGPT agent server logs: which AI browser agents sign requests, which look like Chrome, what you can and cannot identify, and how to allow them safely.
By Crawlito · Updated 7 min read
In short
- OpenAI's crawler page lists GPTBot, OAI-SearchBot, ChatGPT-User and OAI-AdsBot; at the time of writing it does not list ChatGPT agent or Atlas, so a user agent search will not find them.
- OpenAI says ChatGPT agent signs requests with HTTP Message Signatures and sets a Signature-Agent header to https://chatgpt.com; the signature, not the header on its own, is what proves it.
- ChatGPT Atlas used a user agent identical to current Chrome on macOS, and OpenAI scheduled Atlas to stop working on August 9, 2026.
- Google documents Google-Agent as a user-triggered fetcher with a Google-Agent token in its user agent and IP ranges in a published JSON file.
- Allow agents by verifying signatures or published IP ranges at your CDN, never by trusting a user agent or a Signature-Agent header alone.
On this page
- What are AI browser agents and how do they differ from crawlers?
- Why ChatGPT Atlas looks like ordinary Chrome in logs
- How ChatGPT agent signs its requests
- How Google-Agent identifies itself
- What you can detect in analytics and what you cannot
- How to allow legitimate agents through a WAF without opening it to spoofers
- Frequently asked questions
- Check what an AI crawler can reach
You can identify some AI browser agents in your logs and not others. ChatGPT agent signs its requests, so a CDN or your own code can verify it. Google-Agent has a documented user agent token and a published IP range file. ChatGPT Atlas looked like ordinary Chrome by default, so a user agent search cannot separate it from a human visitor. OpenAI scheduled Atlas to stop working on August 9, 2026.
None of this tells you whether an agent will cite or recommend your site. It tells you whether an agent reached it, and whether the visit was real.
What are AI browser agents and how do they differ from crawlers?
A crawler fetches pages on its own schedule. A browser agent opens pages because a person asked it to do a task, such as comparing products or filling in a form, and it may click and scroll like a person would.
| Type | Examples | Trigger | Main way to identify it |
|---|---|---|---|
| Crawler | GPTBot, OAI-SearchBot | The operator's schedule | User agent token plus published IP ranges |
| User-triggered fetcher | ChatGPT-User, Google-Agent | A user's request | User agent token plus published IP ranges |
| Browser agent with signed requests | ChatGPT agent | A user's task | Cryptographic signature on each request |
| Browser agent that looks like a browser | ChatGPT Atlas | A user's task | Behavioral signals only, no reliable token |
The categories overlap, and the last row is the hard one. For the first two, see GPTBot, OAI-SearchBot and ChatGPT-User explained and how to find AI crawlers in your server logs.
Why ChatGPT Atlas looks like ordinary Chrome in logs
OpenAI's crawler documentation lists four agents: OAI-SearchBot, OAI-AdsBot, GPTBot and ChatGPT-User, each with a user agent string and an IP list. At the time of writing, Atlas is not among them, and neither is ChatGPT agent. So there is no OpenAI-documented Atlas token to search for.
What independent observers found:
- Simon Willison reported in October 2025 that Atlas sent a user agent identical to the one he got from current Chrome on macOS (
Chrome/141.0.0.0). - HUMAN Security, a bot-detection vendor, says the default user agent matches Chrome on macOS with no Atlas token, and that simple user agent filters cannot reliably separate Atlas from Chrome. It also reports that some background requests, such as a favicon fetch for a sidebar preview, carried a different user agent naming Atlas. That is one vendor's observation, not OpenAI documentation.
The practical result: an Atlas visit that loads your page can appear in your access log as an ordinary Chrome request.
Atlas was also scheduled for retirement. OpenAI's help center describes deprecating Atlas and moving browser-based agent features into ChatGPT and Codex, with Atlas scheduled to stop working on August 9, 2026 (see also The Register's report). That date has passed, so check OpenAI's page for the current status. Old logs may still hold Atlas visits. For new traffic, look at ChatGPT's current browsing features instead. OpenAI's documentation we could read does not say how those present themselves, so treat that as unknown until it does.
How ChatGPT agent signs its requests
At the time of writing, OpenAI's agent access page calls this the ChatGPT cloud browser, while several CDNs still list it as ChatGPT Agent. OpenAI's allowlisting article says ChatGPT agent signs outbound requests using HTTP Message Signatures (RFC 9421). Each request carries Signature and Signature-Input headers, plus a Signature-Agent header set to "https://chatgpt.com", quotation marks included. This is the Web Bot Auth approach.
Signature-Agent: "https://chatgpt.com"That header alone is just text. Anyone can send it. What proves the origin is a valid signature checked against the public keys OpenAI publishes, which a verifying CDN or your own code can check. The signature scheme is also described in Cloudflare's Web Bot Auth documentation: Signature-Agent must be an https:// URL in double quotes, and Cloudflare checks the signature against the key registered for that agent.
Log the header at your origin
You can record whether the header arrives. In nginx, a request header named Signature-Agent is available as $http_signature_agent:
log_format agents '$remote_addr [$time_local] "$request" $status '
'"$http_user_agent" sig_agent=$http_signature_agent';
access_log /var/log/nginx/agents.log agents;grep 'sig_agent=.*chatgpt.com' /var/log/nginx/agents.log | tail -20Two limits apply. A logged header is a claim, not a verified fact. And if a CDN or proxy sits in front of your origin and strips the signature headers, you will see nothing. OpenAI's article tells site owners to make sure intermediaries preserve Signature, Signature-Input and Signature-Agent.
How Google-Agent identifies itself
Google's Google-Agent documentation describes it as a user-triggered fetcher used by agents hosted on Google infrastructure to navigate the web and perform actions when a user asks. It has two documented user agent strings, one mobile and one desktop, and both contain the token Google-Agent:
Mozilla/5.0 (X11; Linux x86_64) AppleWebKit/537.36 (KHTML, like Gecko; compatible; Google-Agent; +https://developers.google.com/crawling/docs/crawlers-fetchers/google-agent) Chrome/W.X.Y.Z Safari/537.36The Chrome version is a placeholder (W.X.Y.Z) that changes, so match on the Google-Agent token. Google lists user-triggered-agents.json as the IP range file for it, and warns that user agent strings can be spoofed, so check the IP too. Google's page says user-triggered fetchers generally ignore robots.txt rules because a person requested the fetch. It also says Google is experimenting with Web Bot Auth for this agent, using the identity https://agent.bot.goog, without giving signature details. Our Google-Agent guide covers the robots.txt question.
What you can detect in analytics and what you cannot
You can detect:
- Signed or token-bearing requests in server logs, as above.
- Verified agent traffic in your CDN's reports, if your CDN verifies signatures. Cloudflare, for example, treats ChatGPT agent as a signed agent, a category it announced in August 2025 for agents that a user directs, that sign requests with Web Bot Auth and that follow Cloudflare's policy.
You cannot reliably detect:
- Atlas sessions in analytics. A request that presents a Chrome user agent has no token to filter on.
- Intent. A signature says a request came from ChatGPT's cloud browser. It does not say which user started the task or why.
- Whether a visit led to a citation or a purchase. Agent visits and AI citations are different measurements; AI citations vs referral traffic explains why they should be reported separately.
Whether a given agent runs your analytics JavaScript is not documented in the sources used for this post, so do not assume agent visits are absent from, or present in, your analytics numbers. Keep server-log counts and analytics sessions as separate figures, as described in how to measure AI referral traffic and crawler visits.
How to allow legitimate agents through a WAF without opening it to spoofers
Agents that act for users can hit the same bot rules that block scrapers. A WAF challenge page shows an agent a challenge instead of your content. To allow the real ones without allowing impersonators:
- Allow on verification, not on a string. Match on a CDN's verified-agent classification, a verified signature, or the operator's published IP ranges. A rule that allows any request containing
Google-Agentorchatgpt.comin a header lets anyone through by copying it. - Use your CDN's built-in verification where it exists. OpenAI's article says Akamai performs the signature verification for you, that Vercel's bot verification permits ChatGPT agent requests by default, and that on Cloudflare it appears as a signed agent with the tag
chatgpt-agentand detection ID129220581, which a WAF custom rule can match. Detection IDs are available only to Enterprise Bot Management customers. OpenAI also documents a dashboard option: under Security > Settings > AI Crawler and Agent access policies, set "dots, OpenAI" under Personal Agents to Allow. Allowing ChatGPT's cloud browser through a WAF has the per-vendor steps. Cloudflare's own documentation says that as of July 1, 2026 signed agents are part of its Verified bot category, distinguished by a Direct or Intermediary metadata field. Check Cloudflare's current documentation before writing the rule, because the label OpenAI's article gives may be described differently by now. - On other setups, verify the signature yourself. OpenAI's guidance is to confirm that
Signature-Agentexactly matches"https://chatgpt.com"and to verify the signature against the published key. - Keep IP checks for IP-based agents. For Google-Agent, compare the source address with the ranges in Google's file. How to verify GPTBot IPs and spot fake AI crawlers shows the pattern.
- Rate-limit what you allow. Verification says who sent a request, not that the volume is acceptable.
Do not serve agents different content from people. Verification decides whether to let a request through, not what it sees.
Frequently asked questions
Can I find ChatGPT agent in my logs by user agent?
OpenAI's crawler page does not list a user agent for ChatGPT agent. Its allowlisting article describes identification through signed requests and the Signature-Agent header, so check for those, and verify the signature rather than trusting the header.
Is ChatGPT Atlas still around?
OpenAI's help center describes deprecating Atlas, with the browser scheduled to stop working on August 9, 2026, a date that has passed. Atlas visits may still appear in older logs, and OpenAI says the browser-based agent features move into ChatGPT and Codex. Check OpenAI's page for the current status.
Does Google-Agent obey robots.txt?
Google's documentation says user-triggered fetchers generally ignore robots.txt because a user requested the fetch, and lists Google-Agent as a user-triggered fetcher. Use access controls, not robots.txt, if you need to restrict it.
Should I block AI browser agents?
That is a business decision, not a technical default. Blocking removes agents acting for real users who may be trying to buy from or read your site. If you block, block on verified identity, and note that robots.txt may not apply to user-triggered traffic.
Check what an AI crawler can reach
Agents only matter if your pages can be read once they arrive. The free check reads one page the way AI crawlers do, with no account needed, including robots.txt rules, robots meta tags and whether the main text is in the raw HTML. For a whole site, an audit runs the checks across up to 500 pages with evidence and prioritized fixes.
Filed under AI crawlers. Topics: ai-agents, chatgpt, server-logs, web-bot-auth, ai-crawlers