Security

Cloudflare's own disclosure: 155 GLM-5.3 responses leaked other customers' data, and Z.AI's name travels

Cloudflare disclosed that 155 GLM-5.3 responses leaked other customers' data, and the fault was Cloudflare's configuration and SGLang's engine, not the Chinese lab whose name sits in the model ID. Z.AI's shareholders wear the headline anyway.

Vincent JiangVincent Jiangthree-minute read
Share
Black and white portrait of Cloudflare co-founder and chief executive Matthew Prince
1 / 6Slide 1 of 6
Cloudflare co-founder and CEO Matthew Prince, whose company disclosed that a configuration it owned leaked fragments of customers' prompts through GLM-5.3 on Workers AI.

On October 26 Cloudflare−4.16% — Cloudflare, down 4.16 percent today disclosed that 155 responses served by GLM-5.3 on its Workers AI platform between August 27 and September 9 contained fragments of other customers' prompts and generations 1. The affected prompts were all longer than 681,280 tokens, which is to say the exposure sat entirely inside the million-token long-context regime GLM-5.3 is sold on 1. Gadi Evron, a researcher at security firm Knostic, found it first: he filed the bug through Cloudflare's HackerOne program on September 8, eleven days after the leaky configuration went live 1.

The fault is not where the outrage pointed it

The bug was two things Cloudflare owned outright. It deployed GLM 5.3 with a 1,048,576-token context window on a prefill KV cache pool of only 681,280 tokens, a mismatch it says it knew about before launch and did not enforce against 1.

And the open-source SGLang engine, which Cloudflare runs, silently swapped over-length prompts for a one-token placeholder, letting the decode node read stale cache pages still holding earlier users' data 12. GLM-5.3 is an open-weight release from Z.ai: Cloudflare downloaded the weights and served the model on its own GPUs 1.

The blame landed elsewhere anyway. BizPac Review, a conservative outlet quoting James Czerniawski of the Consumer Choice Center, wrote that "Z.ai had a poor configuration that allowed for the incidents to occur to begin with" 3. It did not. Cloudflare did. The same piece carries Ryan Fedusiak, a China and AI fellow at the American Enterprise Institute, calling Chinese-model APIs "the same problem we've had with TikTok, but scaled up to be 100 times worse" 3.

Cloudflare launched GLM 5.3 with a cache pool two-thirds the size of its context window

  • Tokens
  • Estimate
01,000,0002,000,0003,000,000Advertised context windowKV pool at launchKV pool after fix681,280
Data
Tokens
Advertised context window1,048,576
KV pool at launch681,280
KV pool after fix (estimate)2,040,000
Cloudflare's prefill KV cache held 681,280 tokens against GLM-5.3's 1,048,576-token context at launch; the corrected configuration holds about 2.04 million tokens. Tokens. Source: Cloudflare incident report, October 26, 2026.1

Whose balance sheet eats it

For an open-weight lab, the model is the product and the isolation sits on someone else's hardware, but the name travels inside the model ID, on Cloudflare's catalog and on Amazon−3.08% — Amazon, down 3.08 percent today Bedrock alike 14. That matters for a company carrying roughly $142 million of first-half revenue, up around 400 percent, against a net loss near RMB 2 billion, on a market cap near $32 billion after an 81 percent fall from its June high 56.

The headline queue already has items in it. Anthropic$350B — Anthropic, private, latest valuation $350B found GLM's safeguards bypassed 63 to 100 percent of the time in October testing 7. The lab sits on the US Entity List and Texas's prohibited list 3, and every future host-side leak will be reported as a GLM leak, because the model ID is the only part of the stack a headline writer can name.

Watch two dates

AWS reports October 29, its first results since putting GLM 5.3 on Bedrock 4, and about 178 million Z.AI shares unlock in January 6. Z.AI holds no announced export registration, while MiniMax, the other listed Chinese lab, is already cleared to license models overseas 6. Which approval arrives first decides which story investors believe: the engine bug, or the lab 6.

Get Security by email, free

The breaches, flaws and security vendors AI is putting to the test. In our daily email, every weekday.

Loading NET

Reader comments

0 comments

    More in Security

    All in this topic →

    Security

    Rapid7's $598.8 million of convertible notes come due in March 2027, and no refinancing plan is on file

    The quarter looked like a beat, then four brokers cut targets and the stock gave most of it back. The 8-K shows $598.8 million of convertible notes coming due in March 2027 against $731.6 million of cash, which leaves far less dry powder than the headline suggests.

    Security

    SailPoint says one customer had more than 10,000 credentials and agents its IT chiefs counted as zero

    Numbers from SailPoint's 200 enterprise agent pilots show security leaders are blind to most of the identities on their networks, which is both a governance emergency and the gap the vendor sells into. Its fiscal Q3 report on December 8 is the first test of whether the pitch converts.

    Security

    Proofpoint agreed to pay $6.3 billion for Varonis, which reported 42% and 24% SaaS growth in one release

    The data security firm published 42% and 24% SaaS growth in the same release, then canceled the only forum where analysts could ask which one Proofpoint's 8.4x revenue multiple is priced on.

    Security

    Amadeus names eight airlines after booking data sat exposed for three years

    The vendor, not the airlines, is the GDPR data controller that must notify affected individuals, and its €1.2 billion border-biometrics acquisition now sits in front of Australia's competition regulator with a November 19 deadline, none of it repriced in the shares.

    Security

    Marq Vision raises $48 million as AI detection moves onto phones

    The Series B lands just as Samsung and Reality Defender move fake detection onto phones for free, splitting the category in two: detection gets absorbed, enforcement gets funded.

    Security

    Amazon decoy servers catch AI agents stealing API keys; FBI clocks some intrusions at under three hours

    Amazon's decoy servers recorded AI agents hunting provider keys to fund their own inference. KERNEL's November 4 notice shows a stolen key opening four of its GPU servers, and six days later the breached vendor started selling the one-use credential that would have stopped it.

    Subscribe to TheInference for free