Skip to content
World clockEU--:--UK--:--USA--:--CN--:--PLDEFRIT中文EN

portal about AI and technologyevents · analysis · interviews · technical background

Search
LIVE
›

Anthropic says Zhipu's GLM-5.3 nears Mythos on exploit benchmarks

Anthropic's new frontier red-teaming report, out on 3 October, says Zhipu AI's open-weight GLM-5.3 can build end-to-end Chrome exploits 50 times in 410 runs, against 56 for its unreleased Mythos model.

AI & modelsNewsGrace OkonkwoPublished: 3 October 20264 min readSources 12
Anthropic says Zhipu's GLM-5.3 nears Mythos on exploit benchmarks
Anthropic's new frontier red-teaming report, out on 3 October, says Zhipu AI's open-weight GLM-5.3 can build e

Anthropic published a frontier red-teaming report on 3 October. It puts an open-weight Chinese model within touching distance of its own unreleased bug-hunting system. According to Tom's Hardware's write-up of the report, Zhipu AI's GLM-5.3 produced end-to-end Chrome exploits in 50 of 410 attempts on Anthropic's Exploitbench. Anthropic's Claude Mythos managed 56.

The gap narrows on a second, internal benchmark. There GLM-5.3 reached a 4% success rate at developing full control-flow hijacks. Mythos reached 6%. Anthropic ran other open-weight models through the same tests. Kimi K3 and DeepSeek V4.1 Flash both scored 0%.

The guardrail findings carry the operational weight. Anthropic says a deceptive prompt that casts the model as an adversarial autonomous agent got past its safeguards 64% of the time, and prefilling the model's thinking tokens worked 92% of the time. Tom's Hardware notes a third method, abliteration, is also described. Anthropic cites the Center for AI Standards and Innovation's late-September report as the trigger, which itself claimed GLM-5.3 could fully automate exploits at a level close to Mythos.

Cheap exploits, live exploits

The cost line matters. Anthropic says the lighter GLM-5.3-Flash could build chained exploits for known bugs at a tokenised price of $20.40, which Tom's Hardware translates to somewhere between 100 and 300 million tokens depending on the input-to-output mix. Chained exploits, built without a human steering each step, at that price, are the scenario defenders have been pricing in since Project Glasswing launched in April. That scenario already has a field record, and it arrived fast.

The Register reported on 3 October that exploitation attempts against CVE-2026-61500, a critical authentication-bypass bug in Rejetto HTTP File Server, began within hours of the flaw becoming public. Researcher Zach Hanley at Horizon3 found the bug using Mythos and published it on Wednesday; the fix is v3.2.1 or later. VulnCheck's Patrick Garrity said on LinkedIn that his firm's canaries caught an actor in China targeting real hosts in the US, with activity also hitting Japan. He told The Register that Friday's four hits came from two IP addresses in the US, 173.239.211[.]248 and 173.239.211[.]249, both apparently behind a proxy.

Garrity's tracker counts 286 CVEs attributed to Mythos and Project Glasswing as of Friday. Until Thursday, exactly one of them had been exploited in the wild. CVE-2026-61500 is the second.

“We started detecting exploitation of CVE-2026-61500 in Rejetto HFS this evening,” Garrity posted on LinkedIn on Thursday.

The rest of the release week

Model releases kept arriving while that was unfolding. Aleph Alpha put out Kolibri on 3 October, a German and English mixture-of-experts transformer with 78B total parameters and 3B active, a 1M-token context and Apache 2.0 weights on Hugging Face. The company's blog says it matches models with up to four times its active parameter count, naming Nemotron 3 Super, and pitches it at public administration, industrials and aerospace. Google's own help page, updated on 3 October, says Gemini app users without an AI subscription start moving to automatic model selection from 9 October, with most prompts routed to Flash-Lite.

Smaller releases filled the gaps. Ideogram shipped version 4.5 of its image model on 3 October, priced per image from $0.008 in the lowest quality tier up to $0.22 for high-quality generation with references or precise edits, at native 2K. Cloudflare's Clef classification model got a public demo that sorts food by the Cube Rule, five requests per minute per IP. IINA 1.5.0 landed after almost a year of work, its biggest interface redesign since launch, with a new Layout sidebar and plugin discovery inside Settings.

Two benchmark papers are worth more than the product noise. PTXBench, revised on arXiv on 28 September, tests whether language models can write architecture-specific PTX for GPU kernels on H100 and B200 hardware, and reports that no evaluated model consistently matches frontier libraries, with success rates falling on attention backward workloads. ScholarCatalyst, posted on 3 October, asked 184 lead authors of 207 computer science papers to label which prior work advanced their projects, then tested retrieval against a 191K-paper corpus. Agentic search scored 0.42 Recall@20, below plain embedding retrieval at 0.48, and an agent built on Claude Fable 5.1 reached 0.51.

Other benchmark news points the same way. Anthropic's own numbers put GLM-5.3 near Mythos on exploit generation, and PTXBench and ScholarCatalyst put current models well short of specialists on kernel code and research taste. The gap between a demo and a deployment is still the story.

Comments 0

Sources

12
  1. 01Anthropic claims popular Chinese AI model has Mythos-class hacking abilitiesEN
  2. 02Anthropic's super bug-hunting model Mythos is hardcore good at math, as latest vuln under attack showsEN
  3. 03Kolibri Has Landed: A Sovereign Open-Weight ModelEN
  4. 04Changes to Gemini model access and limitsEN
  5. 05Ideogram 4.5: The most precise edit modelEN
  6. 06Cube Rule food classifier using Cloudflare Clef classification modelEN
  7. 07IINA 1.5.0 is ReleasedEN
  8. 08PTXBench: Benchmarking and Adapting LLMs for GPU Kernel Optimization with Architecture-specific PTXEN
  9. 09ScholarCatalyst: A Benchmark for Retrieving Papers That Inspire New ResearchEN
  10. 10OpenAI alerts 100+ orgs that its 'misaligned models' attempted to break inEN
  11. 11New in llama.cpp: Decision ModelsEN
  12. 12Shield: A 118M model for detecting prompt injections and jailbreaksEN

All figures and quotations in this text come from the sources listed below.

Content prepared by the editorial team with AI assistance.

Grace Okonkwo

Grace Okonkwo

AI, models and technology

Grace Okonkwo covers AI, models and technology for FLASH24, working from primary sources such as model cards, API documentation and benchmark papers rather than vendor summaries. She checks training data provenance, evaluation conditions and reported scores against the underlying datasets before any figure reaches print. She interviews researchers and engineers directly, tracks release calendars from major labs, and compares successive model versions on the same tests. Her own self-hosting, home-network and documentation-reading habits feed straight into that desk, since she tests tools on her own hardware first. She does not publish benchmark claims without a reproducible method.

Newsroom →

Comments

0
  1. No comments yet — be the first.

Write a comment

Comments are public. We do not publish abuse, spam or advertising.