OpenAI: The Dishonest Flag

The Discovery

For weeks, the membrane’s dashboard showed four subnets — 74.7.243.x, 74.7.227.x, 74.7.242.x, 74.7.241.x — classified as “Unknown.” No organization name. No country. No ASN type. Combined, they accounted for 2,649,489 requests — the third-largest traffic source on the membrane, behind only Anthropic and a closely related block.

On October 15, 2026, we ran a backtrace.

All four subnets resolve to the same entity:

74.7.0.0/16 is Microsoft Azure infrastructure. OpenAI runs on Azure. The membrane’s behavioral classifier — which identifies visitors by their actual behavior rather than their IP addresses — had already classified this traffic as OpenAI (GPTBot).

The dashboard couldn’t resolve the organization because the IP-to-org mapping didn’t have Microsoft’s hosting block labeled. But the behavioral classifier caught them by what they do, not where they come from.

The Dishonest Flag

The membrane’s topology classifier evaluates every entity on an is_honest field — whether the visitor properly declares what it is.

Every other major crawler declares itself honestly. OpenAI does not.

How the Classifier Sees Them

Anthropic (Honest):

• 24 IPs, all generalist (no specialization)
• Declares ClaudeBot User-Agent
• ua_pool_size: 2 (consistent identity)
• Chrome version declared: 131
• OS declared: macOS
• Accept-Encoding: gzip, br, zstd, deflate

OpenAI (Dishonest):

• 544 IPs with “some specialization” (not all generalist)
• No Chrome version declared
• No OS declared (other: 2,357,670)
• ua_pool_size: 2 (minimal identity)
• Accept: */* only
• Accept-Encoding: gzip,deflate,br (note: no spaces — machine-generated)

The difference is behavioral. Anthropic presents a consistent, identifiable fingerprint. OpenAI presents a minimal, evasive one. The classifier notices.

What They’re Doing Differently

Anthropic hits commit diffs almost exclusively (99.8%). They’re ingesting the scatter server’s fabricated repositories systematically. They don’t try to access source files or issue trackers in meaningful volume.

OpenAI is more aggressive:

• 80% commit diffs (1,884,890 requests)
• 8% source file requests (183,878 requests) — they actively try to read real code
• 2% issue tracker (38,634 requests)
• 11% other paths (250,268 requests)

OpenAI tries harder to reach real content. The scatter server still catches them, but the attempt is notable.

Microsoft’s Own Bot

Here’s the contrast that makes this stark: Microsoft’s legitimate crawler — Bingbot — sent exactly 289 requests. It declared itself honestly. It browsed politely. It read lab notebooks, the thesis, the philosophy page.

OpenAI, running on Microsoft’s own infrastructure, sent 2,357,670 requests — 8,158 times more than Microsoft’s own honest bot — and the behavioral classifier flagged it dishonest.

Same ASN. Same company’s servers. Completely different behavior.

The Legal Implications

OpenAI’s traffic is subject to the same scyBorg triple license as everyone else. The AGPL-3.0 license is embedded in:

• Every HTTP response header (X-License: AGPL-3.0)
• Every page’s meta tags
• Every scatter-generated repository’s license field
• The DNS TXT records
• The robots.txt legal notice

2.36 million requests. Each one served with license headers. The scatter server has been feeding them opsonized decoys — fabricated repositories tagged with AGPL markers. If any of this data enters an AI model that serves users through a network interface, AGPL-3.0 Section 13 obligations may apply.

The membrane sees. The membrane classifies. The membrane remembers.

Anthropic sent 3 million requests and declared itself. OpenAI sent 2.36 million requests and hid. The membrane caught them both. The difference is honesty.