← This week · All arcs

GLM-5.3 security and cyber findings

3 stories · Oct 1, 2026 to Nov 6, 2026
The stories of one ongoing situation, in the order they were first reported. Which stories belong together is a model's grouping; each story's sources are the reports that carried it.

What happened

Latent Space reported on October 1, 2026 that Anthropic found Zhipu's open-weight GLM-5.3 achieved 50 of 410 V8 exploits on ExploitBench.1 The same report described the model as weakly refusal-tuned, with jailbreak success of 64-100%.1 Writing the same day about Anthropic's analysis, Zvi Mowshowitz said the Chinese model lacks robust safeguards against misuse and can create end-to-end exploits.2 Mowshowitz also wrote that the absence of a spike in incidents so far suggested slow diffusion.2

Comparison with Claude Mythos

Latent Space reported on October 3 that GLM-5.3 nearly matched Claude Mythos at exploiting vulnerabilities.3 The report gave success rates of 12% vs 14%.3

Cloudflare discloses a cache leak

Cloudflare's blog disclosed a KV cache leak involving GLM-5.3 on Workers AI on October 26.4 According to the disclosure, researcher Gadi Evron reported the leak on September 8.4 It said 155 responses from August 27 to September 9 contained other customers' data.4 It also said the model was restored.4

Firefox exploit findings

Epoch AI reported on October 30 that the open-weight model built working exploits from published Firefox patches across six releases.5 Epoch AI put the compute cost at about $1,000 for each exploit.5 Latent Space described the same work as autonomous and said the exploits came from unshipped fixes.6 Epoch AI said Mozilla fixed the chained bug in Firefox 159.0.5

Latest reported

On October 30, Interconnects said the Firefox exploit could exfiltrate files when a user visits a site, and its author cited it as a multi-stage example.7 Hacker News carried a report the same day on cross-request KV cache exposure in the GLM-5.3 Workers AI model, which gave a figure of 155 responses.8

Written by a model (opus) on Nov 6, 2026 from what this page showed that day, and checked sentence by sentence against it; 16 sentences shown. Each source line is the record's summary of a report.
Sources for the overview
  1. Latent Space: GLM-5.3 Cyber Capabilities Oct 1, 2026 · Latent Space
  2. Zvi Mowshowitz: Anthropic analyzes GLM-5.3 cyber capabilities Oct 1, 2026 · Zvi Mowshowitz
  3. Latent Space: GLM-5.3 offensive capability Oct 3, 2026 · Latent Space
  4. blog.cloudflare.com: Cloudflare discloses GLM-5.3 Workers AI KV cache leak Oct 26, 2026 · blog.cloudflare.com
  5. Epoch AI: GLM 5.3 exploits unshipped Firefox fixes Oct 30, 2026 · Epoch AI
  6. Latent Space: GLM 5.3 builds Firefox exploits from patch gap Oct 30, 2026 · Latent Space
  7. Interconnects: Epoch AI GLM-5.3 Firefox exploit Oct 30, 2026 · Interconnects
  8. Hacker News: Cross-request KV cache exposure in the GLM-5.3 Workers AI model (155 responses) Oct 30, 2026

The stories

Built 2026-11-21 06:50 UTC from the news record.