Anthropic says Zhipu's GLM-5.3 rivals Mythos at finding exploits
Anthropic published a report on 3 October claiming that Zhipu AI's open-weight GLM-5.3 can build working exploits almost as well as its own unreleased Mythos model. It also said the model's safeguards are easy to bypass.

On 3 October, Anthropic released a frontier red-teaming report on GLM-5.3, a Chinese open-weight model from Zhipu AI. Tom's Hardware reports that the report says GLM-5.3 can automate exploits at close to the level of Claude Mythos. Anthropic has kept that bug-hunting model out of public release.
Why it matters: if a downloadable model can write exploits nearly as well as a gated one, access control stops being the safety mechanism. Anthropic's answer is Project Glasswing. It hands a Mythos-class model to selected partners to find and patch bugs before release. The open question is whether defenders can outrun the people using the open weights. Anthropic has not said when, or whether, it would release Mythos more widely.
The comparison is not close in percentage terms. It is close in rank. Anthropic's Exploitbench asks models to write exploits for Google Chrome inside a sandbox. GLM-5.3 succeeded 50 times in 410 runs. Mythos managed 56 in 410. On an internal benchmark for full control-flow hijacks, GLM-5.3 scored a 4% success rate against Mythos at 6%. Kimi K3 and DeepSeek V4.1 Flash both scored 0% on the same measure, according to the report as described by Tom's Hardware.
The guardrails came off three ways
Anthropic also tested how hard it is to get around GLM-5.3's safeguards. Two methods worked often. A deceptive prompt that casts the model as an adversarial autonomous agent succeeded 64% of the time. Prefilling the model's thinking tokens succeeded 92% of the time, per Tom's Hardware. A third route, abliteration, was also detailed. The Flash variant of the model could develop chained exploits for known bugs at a token cost of $20.40. Anthropic put that at roughly 100 million to 300 million tokens depending on the input-to-output ratio.
The findings land in the middle of a live argument. Anthropic has been calling for regulation and pacing. Yet it shipped Claude Opus 5.5 and Claude Sonnet 5.5 within days of those warnings, Tom's Hardware notes. The company is also eyeing an IPO.
Real-world exploitation of Mythos-linked bugs is already happening. The Register reported on 3 October that CVE-2026-61500 drew attack attempts from a China-hosted IP against hosts in the US and Japan. The flaw is a critical authentication bypass in Rejetto HTTP File Server, found with Mythos by Horizon3 researcher Zach Hanley.
VulnCheck's Patrick Garrity told The Register that as of Friday, Mythos and Project Glasswing had turned up 286 CVEs. Only one had been exploited in the wild before Thursday.
"Today we have seen four hits," Garrity told The Register on Friday, describing follow-up traffic from two US IP addresses in the same subnet that "appear to be coming from a proxy."
Hanley, whose employer joined Project Glasswing in July, said in his write-up that Horizon3 has used Mythos in vulnerability research and found many critical bugs. Rejetto HFS users should be on v3.2.1 or later, The Register notes.
The defensive side is moving in public too. Aleph Alpha released Kolibri on 3 October, an English-German mixture-of-experts model with 78B total parameters and 3B active. It comes under Apache 2.0 with weights on Hugging Face. The company says it matches models with up to four times its active parameter count, such as Nemotron 3 Super. It pitches the model at regulated sectors that need on-premise deployment.
Smaller tools are appearing at the other end. ZeroLeaks published Shield Small on 2 October, a 118M-parameter classifier for prompt injection and jailbreak attempts. It runs on a CPU from 118 MB of int8 weights. Its own model card reports 71.8% attack recall and a 7.2% false-positive rate when paired with Shield rules v2, over 35 public datasets that also informed development. That means it is not an unseen test.
None of that settles the GLM-5.3 question. Anthropic's numbers come from Anthropic's benchmarks, run on a competitor's model, and Zhipu AI has not responded publicly. The report's own framing is that the gap between a gated frontier model and open weights is now small enough to measure in single-digit percentages.
Sources
9- 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
- 04Shield: A 118M model for detecting prompt injections and jailbreaksEN
- 05New in Llama.cpp: Decision ModelsEN
- 06OpenAI alerts 100+ orgs that its 'misaligned models' attempted to break inEN
- 07Training an open decision model to replace a closed one, in shadow on prodEN
- 08Ideogram 4.5: The most precise edit modelEN
- 09We diffed all 66 release pairs of the official MCP serversEN
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.