Anthropic puts numbers on AI hacking: GLM-5.3 scores 50 exploits to Mythos's 56
Anthropic's new frontier red-teaming report claims Zhipu AI's open-weight GLM-5.3 can build end-to-end exploits on 50 of 410 runs, close behind its own unreleased Mythos model at 56, according to Tom's Hardware.

Anthropic published a report this week arguing that a popular Chinese open-weight model can do much of what its restricted bug-hunting model can do. Tom's Hardware reported on 3 October that the company tested Zhipu AI's GLM-5.3 in a sandboxed benchmark called Exploitbench, where models try to build working exploits for Google Chrome.
GLM-5.3 succeeded 50 times in 410 runs. Anthropic's Claude Mythos managed 56 in the same number of attempts. The gap is seven percentage points wide at most. The report puts a number on a worry regulators have raised for months: capability that Anthropic keeps behind a closed door is already downloadable.
The same tests suggest the distance is not uniform. On an internal benchmark for "full control-flow hijacks", GLM-5.3 scored a 4% success rate against Mythos at 6%. Two other open-weight models, Kimi K3 and DeepSeek V4.1 Flash, scored 0% on that measure, according to the same report.
Safeguards are the other half of the claim. Tom's Hardware reports that Anthropic dodged GLM-5.3's guardrails with a deceptive prompt in 64% of attempts, and with prefilled thinking tokens in 92%. A third method, abliteration, is also described. Anthropic cites a late September report from the Center for AI Standards and Innovation, which it says found GLM-5.3 could fully automate exploits at a level similar to Mythos.
Cost is where the story gets concrete. Anthropic says the lighter GLM-5.3-Flash built chained exploits on known bugs for $20.40 in tokens. Chained exploits string several flaws together, and they are the difference between a proof of concept and an intrusion.
Why the Mythos comparison matters
Mythos is not generally available. Anthropic gives select partners access through Project Glasswing, an initiative it announced in April. The pitch is that defenders get the model first, so bugs are patched before attackers have the same tool.
The numbers above undercut the timing. If a $20.40 open model can chain exploits, the window between discovery and abuse is short. Neither Anthropic nor Zhipu AI has said how the report will change access rules, and Zhipu AI has not responded publicly to the specific benchmark claims.
The defensive side has its own evidence. The Register reported on 3 October that a critical authentication-bypass bug in Rejetto HTTP File Server, tracked as CVE-2026-61500, was found with Mythos and then exploited in the wild within days. Horizon3 researcher Zach Hanley, who found the flaw, published an exploit video. Users should update to v3.2.1 or later.
"We started detecting exploitation of CVE-2026-61500 in Rejetto HFS this evening," VulnCheck researcher Patrick Garrity wrote on LinkedIn on Thursday, per The Register.
Garrity told The Register the Thursday activity came from one IP address in China and targeted vulnerable hosts in the US and Japan. By Friday he had seen four hits from two US addresses in the same subnet, which he said appeared to be a proxy. His tracker counts 286 CVEs found by Mythos and Project Glasswing; until that Thursday, only one had been exploited.
That is the benchmark that matters more than any score. A model that finds bugs faster than vendors patch them shifts risk onto anyone running unpatched software, and HFS had already appeared on CISA's Known Exploited Vulnerabilities catalog in 2024.
Reasoning is not the same as deciding
The measurement problem runs through the rest of the week's AI releases. Aleph Alpha published Kolibri on 3 October, a 78B-parameter mixture-of-experts model with 3B active parameters, a 1M-token context and Apache 2.0 weights. The company says it matches models with up to four times its active parameter count, such as Nemotron 3 Super, and admits public benchmarks miss sector needs, so it built internal ones.
Smaller tools aim at narrower questions. The llama.cpp project added a decision-model endpoint on 2 October, per a Hugging Face community post by Xuan-Son Nguyen and Victor Mustar. It returns a probability for each option in one forward pass. ZeroLeaks released Shield Small, a 118M-parameter classifier for prompt injection, reporting 84.3% mean category-balanced accuracy with its rules, a figure its own card calls a known test set rather than an unseen evaluation.
Research benchmarks show the same unevenness. A paper on arXiv, revised on 28 September, introduces PTXBench for GPU kernel work. No evaluated model consistently matched frontier libraries, and success fell on complex attention backward workloads. ScholarCatalyst, a retrieval benchmark built from 184 lead authors and 207 papers, found agentic search scoring 0.42 Recall@20 against 0.48 for plain embedding retrieval.
Those results are a useful counterweight to exploit counts. A model can look strong on one task and ordinary on the next, and the benchmark you pick decides the headline. Anthropic's report has not been independently reproduced, and the Exploitbench setup is the company's own.
Sources
8- 01Anthropic claims popular Chinese AI model has Mythos-class hacking abilitiesEN
- 02Anthropic's super bug-hunting model Mythos is hardcore good at math, as latest vuln under attack showsEN
- 03Kolibri Has Landed: A Sovereign Open-Weight ModelEN
- 04New in Llama.cpp: Decision ModelsEN
- 05Shield: A 118M model for detecting prompt injections and jailbreaksEN
- 06PTXBench: Benchmarking and Adapting LLMs for GPU Kernel OptimizationEN
- 07ScholarCatalyst: A Benchmark for Retrieving Papers That Inspire New ResearchEN
- 08What does the advent of powerful AI models mean for mathematicians like me?EN
All figures and quotations in this text come from the sources listed below.
Content prepared by the editorial team with AI assistance.
Comments
0- No comments yet — be the first.