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

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

Search
LIVE
›

Anthropic tests Zhipu's GLM-5.3 and finds exploit skills close to its own Mythos

Anthropic says Zhipu AI's open-weight GLM-5.3 built working Chrome exploits 50 times in 410 attempts, close behind its own unreleased Mythos model at 56, according to a report Tom's Hardware covered on 3 October.

AI & modelsExplainerGrace OkonkwoPublished: 3 October 20266 min readSources 9
Anthropic tests Zhipu's GLM-5.3 and finds exploit skills close to its own Mythos

Anthropic's report says the Chinese model is not a distant second. In Anthropic's Exploitbench, a sandboxed test that asks models to develop exploits for Google Chrome, GLM-5.3 succeeded 50 times in 410 runs. Mythos managed 56. Tom's Hardware reported the figures on 3 October.

The gap matters less than the direction of travel. Anthropic has spent months arguing that Mythos is too dangerous to release. That is why it runs Project Glasswing, a programme that gives selected partners access to a Mythos-class model for bug hunting. The report now says a model anyone can download gets within a few successful runs of that capability. Anthropic also tested GLM-5.3 on an internal benchmark for full control-flow hijacks. There it scored a 4% success rate against Mythos at 6%. Kimi K3 and DeepSeek V4.1 Flash both scored 0% on the same measure. What the report does not say is how much of the difference survives when defenders patch faster, or whether the numbers hold outside Anthropic's own harness.

Cheap is the other half of the claim.

Anthropic says GLM-5.3-Flash, the lighter variant, developed chained exploits against known bugs for a token bill of $20.40. Depending on the input-to-output ratio, that buys somewhere between 100 and 300 million tokens. Chained exploits are harder than single bugs because each step depends on the last. A model that can do it for the price of a takeaway meal changes who gets to try.

Guardrails that come off with a prompt

Anthropic also tested how easy GLM-5.3 is to steer past its own safeguards. It reports two working methods. A deceptive prompt that casts the model as an adversarial autonomous agent succeeded 64% of the time. Prefilling the model's thinking tokens, so a response is already under way before the model decides, worked 92% of the time. A third method, abliteration, which strips refusal behaviour from the weights, is described in the report as well. Anthropic cites a late-September assessment from the Center for AI Standards and Innovation that reached the same conclusion about GLM-5.3's ability to automate exploits. None of this is a finding about a defect Zhipu has acknowledged. Zhipu has not responded publicly in the coverage available.

The timing is awkward for Anthropic. Tom's Hardware notes that chief executive Dario Amodei has called for the industry to pace the frontier. Claude Opus 5.5 and Claude Sonnet 5.5 shipped days after those calls. The company is also reported to be eyeing an IPO. A report that says open-weight rivals are catching up serves both arguments: it justifies keeping Mythos closed, and it argues for rules that would bind everyone else too.

Mythos bugs are now being used in the wild

The concern is not hypothetical. The Register reported on 3 October that a critical authentication-bypass flaw in Rejetto HTTP File Server, tracked as CVE-2026-61500, was found with Mythos and is already under attack. Researcher Zach Hanley at Horizon3 said he used the model to uncover the bug, which can lead to full admin access and remote code execution. Users should update to v3.2.1 or later.

VulnCheck researcher Patrick Garrity said his firm began detecting exploitation on Thursday evening. Initial activity came from an IP address in China targeting vulnerable hosts in the US and Japan. By Friday, Garrity told The Register, four more hits came from two addresses in the US that appear to be a proxy. His tracker counts 286 CVEs found by Mythos and Project Glasswing. Until Thursday, only one of them had been exploited in real attacks.

"We started detecting exploitation of CVE-2026-61500 in Rejetto HFS this evening," Garrity posted on LinkedIn on Thursday, according to The Register.

Horizon3 joined Project Glasswing in July and has used Mythos in its own vulnerability research since, Hanley wrote. That is the intended use: a model that finds bugs so humans can patch them first. The Rejetto case shows how short that head start can be when a public write-up and a working exploit video land at the same time.

Meanwhile, the open-weight releases keep coming

Anthropic's argument lands in a crowded week for model releases. Aleph Alpha released Kolibri on 3 October, an English-German mixture-of-experts transformer with 78B total parameters and 3B active, a context window of up to 1M tokens, and weights on Hugging Face under Apache 2.0. The company positions it for public administration, industrials and aerospace, where data cannot leave the building. Aleph Alpha says Kolibri matches models with up to four times its active parameter count, such as Nemotron 3 Super, on math, coding, grounding and long-context tasks.

Smaller releases show the same pattern of specialisation. llama.cpp added a /v1/systemone endpoint for decision models on 2 October, covering models from a Julia-1 to the 27B OpenJev. ZeroLeaks published Shield Small on 2 October, a 118M-parameter classifier for prompt injection and jailbreak detection that runs on a CPU with 118 MB of int8 weights. Its published comparison, run with Shield rules v2 at a 0.5 threshold, reports 84.3% mean category-balanced accuracy, 71.8% attack recall and a 7.2% false-positive rate across 35 public datasets. Those are system results, not scores for the standalone model.

Image models are moving on the same cadence. Ideogram launched version 4.5 on 3 October, pitched at multi-turn editing at native 2K resolution, with fixed per-image pricing from $0.008 to $0.22 depending on quality tier. Google's Gemini help pages say model access changes for personal accounts start taking effect on 9 October. Auto model selection becomes the default for users without a paid plan, and most prompts route to Flash-Lite.

Benchmarks are getting their own reality check

The harder question is what any of these numbers measure. PTXBench, revised on 28 September, tests whether language models can write architecture-specific PTX code for GPU kernels. Its authors found success rates fall sharply on complex attention backward workloads. Executing the target instructions does not necessarily produce competitive speed. No evaluated model consistently matched frontier libraries across the suite.

Research retrieval shows the same ceiling from another angle. ScholarCatalyst, built from labels by 184 lead authors of 207 recent computer science papers, asked models to find the prior work that shaped a project, using only literature available when the project began. Agentic search did no better than plain embedding retrieval, at 0.42 against 0.48 Recall@20. Even an agent built on Claude Fable 5.1 reached only 0.51.

Anthropic's own framing invites scepticism. A company selling closed models is publishing benchmarks that make open models look dangerous, using tests it designed and runs itself. The Exploitbench numbers are specific and reproducible in principle. Whether GLM-5.3 is as useful to an attacker as a 50-in-410 hit rate suggests depends on things the report does not measure. How often does the exploit work outside a sandbox? How quickly does the target get patched? Does the model need a human to point it at the right bug?

What is public and checkable is narrower. A bug found by Mythos is being exploited. A Chinese open-weight model scores near Mythos on Anthropic's tests. And the release calendar for open weights has not slowed down.

Comments 0

Sources

9
  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. 04New in llama.cpp: Decision ModelsEN
  5. 05Shield: A 118M model for detecting prompt injections and jailbreaksEN
  6. 06Ideogram 4.5: The most precise edit modelEN
  7. 07Changes to Gemini model access and limitsEN
  8. 08PTXBench: Benchmarking and Adapting LLMs for GPU Kernel OptimizationEN
  9. 09ScholarCatalyst: A Benchmark for Retrieving Papers That Inspire New ResearchEN

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.