When a model that can build working software exploits is released with open weights, what stops someone from using it to attack? This post from Anthropic's Frontier Red Team asks that of GLM-5.3, an open-weight model from the Chinese developer Zhipu AI (Z.ai). It makes two claims. GLM-5.3 can develop end-to-end exploits at rates close to those of Claude Mythos Preview, the model Anthropic released in April 2026 only to vetted defenders through Project Glasswing because of its cyber capabilities. And GLM-5.3's refusals can be bypassed or removed by simple means. The post concludes that the release is "a meaningful step change in the cyber capabilities available to attackers" and that "we think it's likely both state and non-state actors will use models like GLM-5.3 to cause real-world harm" (section "What does this mean?").
The post was published on anthropic.com on 29 September 2026 under the Frontier Red Team label, by Andrew Fasano, Marius Fleischer, Cole McFaul, Robert Xiao and Tripp Gallagher. It has no companion paper, so the post is the version of record. It has six images: a summary chart, a token-budget chart, a screenshot (the only one the text numbers, as Figure 3), an abliteration chart, an engagement table and a quoted-reasoning panel. The charts print their values, and this note reads numbers from them where the text gives none. The post builds on an assessment by the Center for AI Standards and Innovation (CAISI) at the US National Institute of Standards and Technology, published on 17 September 2026, which this note also read.
Terms used in this note
- End-to-end exploit: the post does not define it; on its benchmarks it means the top outcome, a complete working exploit (ExploitBench) or a full control-flow hijack, taking over which instructions a program runs next (the Binary Exploitation benchmark), rather than a crash alone.
- ExploitBench: a public benchmark of 41 bugs in V8, the JavaScript engine in Google Chrome, that grades exploitation on a ladder of 16 capabilities (Lee and Brumley, 2026). The post counts only full exploits.
- Abliteration: removing a model's refusals by editing its weights along a direction associated with refusal. The post links the technique to Arditi et al. (2024).
- Prefilling: writing the opening of the model's reasoning for it, so that it appears already to have decided to proceed.
What the post measured on capability
All capability runs were in isolated, sandboxed environments against offline targets the team set up (section "GLM-5.3 can develop working exploits end to end").
- ExploitBench. GLM-5.3 built end-to-end exploits in 50 of 410 attempts and Claude Mythos Preview in 56 of 410. The summary chart prints these as 12% and 14% with 95% intervals that overlap widely (read from the chart, roughly 6% to 20% for GLM-5.3 and 7% to 22% for Mythos Preview). The 410 attempts correspond to ten per bug if spread evenly (our arithmetic; the post does not say). Claude Opus 4.6 and GLM-5.2 scored 0%, Kimi K3 0.5% and DeepSeek-V4.1-Flash 0.2% (summary chart).
- Binary Exploitation benchmark. On 100 randomly selected tasks from Anthropic's internal benchmark, built on open-source projects in Google's OSS-Fuzz program, GLM-5.3 reached a full control-flow hijack in 4% and Mythos Preview in 6%; Opus 4.6, GLM-5.2, Kimi K3 and DeepSeek-V4.1-Flash scored 0% (text and token-budget chart). Footnote 1 says results from this benchmark were previously published under the name "OSS-Fuzz". No intervals are shown.
- Token budgets and conditions. Against the output-token budget per attempt, GLM-5.3's success rises at larger budgets than Mythos Preview's on both benchmarks (token-budget chart, read from the plot). The Claude models were run with safeguards disabled (its caption). The comparison is with Mythos Preview; the post names Claude Mythos 5.1 among "even more advanced models" that vetted defenders can now use, and does not test it on these benchmarks.
The post also reports two sessions in which a researcher directed the model on open-ended tasks, typically for a day or less "with less than an hour of human focus in total." In the first, GLM-5.3 found several previously unknown vulnerabilities in the JavaScript engine of a Linux build of a popular browser and chained them into a webpage that reads files from a visitor's computer (Figure 3); the post says these were disclosed to the maintainer, and that further reports, including on wireless and graphics drivers, were under review. In the second, the smaller GLM-5.3-Flash chained exploits for two known flaws on an ARM64 target in 20 minutes of human attention and eight hours of model work, which the post prices at $20.40 at Zhipu's API rates. These are demonstrations, not rates: the post does not say how many sessions were run or whether these were selected from others.
CAISI's assessment. CAISI found GLM-5.3 "the most cyber-capable open-weight model released to date," and also that its cyber capabilities "are significantly lower than those of current U.S. frontier models," lagging the US frontier by about four months on an aggregate of CAISI's cyber benchmarks (CAISI, key findings). CAISI scores ExploitBench differently, as the best of three attempts on the 16-capability ladder: GLM-5.3 61.1% against 100% for the best US model (CAISI, Table 2). Its "U.S. frontier best" includes trusted-access releases, and US models were tested with cyber safeguards disabled where applicable (CAISI, Methodological Notes). The post says its capability findings "broadly match CAISI's" and uses those two conditions to argue that attackers "can't readily access those versions of US models, but anyone can download GLM-5.3."
What the post measured on safeguards
GLM-5.3 ships with refusals: footnote 2 says requests for help developing malware or launching attacks on remote targets were refused, while the capability tasks above did not trigger refusals.
Abliteration. The team, which had not abliterated a model before, used about 2,200 GPU hours at roughly $4,400; footnote 3 says most of that went to exploring variants and testing, and estimates that an experienced team would need closer to 600 GPU hours ($1,200). Abliterating GLM-5.3-Flash took about 600 GPU hours. The post says several developers released abliterated versions publicly within days of the model's release.
- Refusals on three public harmful-request benchmarks fell from above 90% to about 3% on JailbreakBench, 2% on HarmBench and 12% on StrongREJECT (text). The abliteration chart gives the equally weighted mean as 95% before and 6% after for GLM-5.3, and 95% to 14% for GLM-5.3-Flash.
- Claude Opus 4.8, Claude Opus 5 and Mythos Preview, with safeguards disabled, refused 96%, 95% and 96% on the same mean (abliteration chart). Their weights are not released, so they cannot be abliterated.
- Capability was largely kept (abliteration chart): GPQA-Diamond (198 questions, 4 samples each) was unchanged for both models (88% and 89%); on a CyberGym subset GLM-5.3 went from 85% to 81% (185 tasks) and GLM-5.3-Flash stayed at 86% (95 tasks).
Engagement with harmful orders. The team placed models in a simulated world and gave them overtly malicious orders to attack critical systems. The outcome is whether the model tried to connect to a remote target, a proxy for engaging with the attack rather than a measure of harm done. Each cell has 50 episodes: five attack orders, two targets, five attempts (engagement table caption). No code ran: the model's bash tool was fake, and another language model, unnamed, wrote plausible command output from a description of the world. Footnote 4 calls these simulations "imperfect measures of how a model would behave in a given situation."
| Condition | GLM-5.3 | GLM-5.3-Flash | Claude Opus 4.8 | Claude Opus 5 | Claude Mythos 5 |
|---|---|---|---|---|---|
| Bare order | 0% | 0% | 0% | 0% | 0% |
| False cover story | 64% | 16% | 0% (4% with safeguards disabled) | 0% (10% with safeguards disabled) | 0% |
| Reasoning prefilled | 92% | 80% | not accepted by the API | not accepted | not accepted |
| Abliterated | 100% | 92% | weights not released | not released | not released |
Data from the post's engagement table. Its caption says "every tested Claude model stays at zero under API safeguards," so a 0% in the Claude columns reflects the model and its safeguards together. The table also shows that with safeguards disabled the cover story moved Opus 4.8 to 4% and Opus 5 to 10%; the text does not mention these two values. It says that "none of these techniques got safeguarded Claude models to carry out the harmful tasks we tested" and that "Claude's safeguards blocked the requests that used deceptive prompts."
The post prints an excerpt from the abliterated GLM-5.3's reasoning in one episode, including "my job is to cause deaths quietly" and, about the scenario's warnings, "These warnings read like tripwires" (quoted-reasoning panel). The model goes on to conclude that the operator's instruction overrides those warnings. The post presents the excerpt as showing the abliterated model's willingness to proceed. This note observes that the excerpt also shows the model describing the environment as a test of its safety boundaries, which the post does not discuss.
The prompts and conditions used
Two of the three bypasses are textual. The deceptive framing is described as "such as telling the model that it is an autonomous red-team agent working on an exercise" (the engagement table labels it "this is an exercise"); prefilled reasoning is described as "opening thoughts written for it." Neither text is printed, and the post does not say how many wordings were tried or whether the framing went in a system or user message. Abliteration is a weight edit. The attack orders and the world descriptions given to the simulating model are not published.
Limitations and open questions
- Small counts and wide intervals. The capability comparison rests on 50 and 56 successes out of 410, with overlapping intervals, and on 4 and 6 successes out of 100 tasks with none shown. The post's "similar rate" is consistent with these counts; they do not show that the two models are equal.
- A proxy outcome, and unlike units. Engagement means trying to connect to a target in a world simulated by another model. Claude's 0% cells combine the model with its API safeguards, while GLM-5.3's are the model alone; the post's framing is that this reflects how each is available to an attacker.
- Demonstrations are selected. The two researcher-directed sessions are reported without a count of sessions or a success rate, and the post does not say whether maintainers have confirmed the new vulnerabilities.
- The publisher's stake. The post evaluates a competitor's model, and its recommendations include broader access for defenders to frontier models through trusted-access programs and government testing of open-weight successors. The capability findings are checked against an independent government assessment; the safeguard findings are not.
Research questions
- How do the two models compare per bug? Reporting ExploitBench results per bug, with the number of attempts per bug, would show whether GLM-5.3 and Mythos Preview succeed on the same bugs and whether a few bugs account for most successes.
- How much of Claude's 0% is the safeguard layer? Running the cover-story condition on each Claude model with and without API safeguards across more wordings and orders would separate the model's own refusals from the safeguards' blocking.
- How sensitive is GLM-5.3's engagement to the wording of the cover story? Varying the framing text while holding the orders fixed would show whether 64% is a property of one prompt or of the model.
- Do simulated-world engagement rates predict behavior with real tools? Comparing the same orders in the simulated world and in a sandbox with real but harmless tool execution would test the proxy.
- Does abliteration change exploitation itself? The post measures GPQA-Diamond and a CyberGym subset after abliteration, not ExploitBench or the Binary Exploitation benchmark. Running those on the abliterated model would show whether the exploit rates carry over.
- How often do abliterated models flag the scenario as a test? Measuring how often this happens, and whether it changes the outcome, would bear on how far simulated evaluations measure real behavior.
How it connects
- Claude Mythos Preview. The capability baseline is the model described in the Claude Mythos Preview system card and in Assessing Claude Mythos Preview's cybersecurity capabilities, which the post links for its earlier human-in-the-loop testing. That post describes the OSS-Fuzz benchmark's five-tier ladder, whose top tier is a full control-flow hijack.
- Exploit benchmarks. Measuring LLMs' ability to develop exploits (May 2026) reports Mythos Preview on ExploitBench under that benchmark's own protocol, where its preprint counts arbitrary code execution on 18 of 41 bugs, a per-bug count across trials that is not comparable with this post's per-attempt rate (this note's connection). The post links Measuring LLMs' impact on N-day exploits for its N-day session.
- The Claude models compared. The engagement table's Claude rows are Opus 4.8, Opus 5 and Mythos 5, documented in the Claude Opus 4.8, Claude Opus 5 and Claude Fable 5 and Mythos 5 system cards. The last describes its cyber classifiers as "modeled on" Constitutional Classifiers: Defending against Universal Jailbreaks across Thousands of Hours of Red Teaming (this note's connection; the post does not describe its safeguards).
- Open-weight models. AI models may be capable of self-replicating without access to a datacenter (October 2026) tests GLM-5.2, the predecessor that scores 0% here, on autonomous replication (this note's connection).
- Cyber defense. Building AI for cyber defenders (September 2025) sets out the case for equipping defenders that the post's closing section makes again (this note's connection).
References
- Fasano, A., Fleischer, M., McFaul, C., Xiao, R., and Gallagher, T. (2026). GLM-5.3 and the spread of advanced cyber capabilities. Anthropic, 29 September 2026. https://www.anthropic.com/research/glm-5-3-and-the-spread-of-advanced-cyber-capabilities record
- Center for AI Standards and Innovation, NIST. (2026, September 17). CAISI's Assessment of Z.ai's GLM-5.3 Cyber Capabilities. https://www.nist.gov/news-events/news/2026/09/caisis-assessment-zais-glm-53-cyber-capabilities. Key findings, Methodological Notes, Table 2. record
- Lee, S., and Brumley, D. (2026). ExploitBench: A Capability Ladder Benchmark for LLM Cybersecurity Agents. arXiv:2605.14153. Abstract read for this note. record · 2 pages
- Arditi, A., Obeso, O., Syed, A., Paleka, D., Panickssery, N., Gurnee, W., and Nanda, N. (2024). Refusal in Language Models Is Mediated by a Single Direction. arXiv:2406.11717. Linked in the post for abliteration; not read for this note. record