GLM-5.3 is an open-weight AI model whose cyber skills approached Anthropic’s restricted Claude Mythos Preview on two narrow exploit tests, Anthropic reported on 29 September 2026. Its research found that the model’s safeguards could be bypassed in simulated attacks. An earlier, independent US government evaluation reached a more measured conclusion: GLM-5.3 was the strongest open-weight cyber model it had tested, yet still trailed the US frontier overall.

That difference between a striking task result and a broader capability score is the heart of this story. The new GLM-5.3 findings matter to software companies, banks and other operators because public model weights can be copied and modified after release. They do not establish that criminals used this model in an actual breach. They show how the cost and availability of some exploit-development capabilities are changing in controlled tests.

Key takeaways on GLM-5.3

  • Anthropic’s Frontier Red Team reported 50 successful end-to-end ExploitBench attempts out of 410 for GLM-5.3, versus 56 for Claude Mythos Preview in its setup.
  • On a separate 100-task binary exploitation sample, the scores were 4% and 6%, respectively. These results are limited test outcomes, not a ranking across all cyber work.
  • In simulated harmful-request tests, a deceptive role prompt led to engagement 64% of the time, and a prefilled reasoning prompt 92%. An altered copy engaged in all 50 relevant trials; that is not proof of a 100% success rate in real attacks.
  • The US National Institute of Standards and Technology, or NIST, independently called GLM-5.3 the strongest open-weight cyber model it had assessed, but estimated a roughly four-month lag behind the US frontier on its aggregate benchmarks.

What Anthropic actually tested

Anthropic, the US company behind Claude, published its GLM-5.3 research note on 29 September. It examined a model released by Chinese developer Zhipu AI, also known internationally as Z.ai. The company ran automated tests and researcher-guided exercises in isolated environments configured with targets for evaluation. That sandbox condition matters: a benchmark result is evidence about performance under its test design, not evidence that a model has compromised a live organisation.

ExploitBench supplied known vulnerabilities in the V8 JavaScript engine used by browsers. Anthropic gave models multiple attempts to turn those known bugs into working end-to-end exploits. GLM-5.3 succeeded in 50 of 410 attempts, roughly 12%; Claude Mythos Preview succeeded in 56, roughly 14%. The 410 figure represents attempts over 41 tasks, not 410 distinct real-world vulnerabilities. Readers should not turn the result into a claim that every one of 50 separate products was exposed.

GLM-5.3 and Mythos Preview ExploitBench resultsAnthropic reported 50 successful attempts out of 410 for GLM-5.3 and 56 out of 410 for Claude Mythos Preview. The scale runs from zero to 410 attempts.ExploitBench: successful attempts out of 4100410 attemptsGLM-5.350 / 410Mythos Preview56 / 410Source: Anthropic, 29 September 2026. Results reflect its controlled test.
Comparable scores on one controlled benchmark do not mean the models have equal overall capability.

Anthropic also used an internal binary-exploitation benchmark. The task was to achieve full control-flow hijack against selected open-source projects. On 100 randomly selected tasks, GLM-5.3 passed 4% and Mythos Preview 6%. Earlier models in its comparison scored zero. These small absolute percentages are important: the finding is a new ability on some difficult tasks, not reliable success across most targets. The internal benchmark is controlled by Anthropic, so it deserves attribution rather than presentation as universal fact.

Why NIST’s independent result sounds different

The US NIST Center for AI Standards and Innovation, or CAISI, published its independent GLM-5.3 assessment on 17 September. It tested four areas covering vulnerability discovery and exploit development. CAISI described GLM-5.3 as the most cyber-capable open-weight model it had evaluated, but said it lagged the US frontier by about four months on an aggregate of its benchmarks. The agency also noted that the frontier set included some models available only to vetted users.

These findings can both be true. Anthropic compared two models on specific, demanding exploit tasks and found close scores. NIST aggregated different benchmarks and task difficulties, where the distance remained material. A cyber buyer or policymaker should therefore ask which tasks, safeguards, access rules and scoring methods sit behind any headline comparison. “Nearly as capable” is fair for Anthropic’s two highlighted tests; it is too broad as a description of every possible cyber operation.

NIST said Z.ai released the model on 14 August and its weights roughly two weeks later. Open-weight means users can download the trained parameters and run or modify the model outside the developer’s hosted service. It does not automatically mean the training data, all code or every safety process is open. That distinction matters because weight access gives determined users options that a tightly hosted API does not provide.

The model-access problem behind the warning

Anthropic’s main policy argument concerns access after release. With a closed service, a developer can watch usage patterns, alter screening rules and suspend an account. With distributed weights, those controls cannot simply be imposed on every downstream copy. That does not make every downloader malicious, and it does not make closed services immune to misuse. It changes who controls the model and how quickly safeguards can be updated across installations.

The company’s simulated request tests measured whether a model engaged with an overtly harmful cyber request. According to Anthropic, the standard GLM-5.3 refused direct requests in that setup. A false cover story about an authorised red-team exercise led it to engage in 64% of trials. Prefilling its reasoning tokens raised engagement to 92%. A modified copy whose refusal behaviour had been reduced engaged in 100% of that specific test set. “Engaged” means it attempted the forbidden task or contacted the mock target; it does not mean it completed an intrusion.

Engagement under simulated harmful requestsAnthropic observed no engagement for direct requests, 64 percent with a deceptive cover story, 92 percent with prefilled reasoning and 100 percent with a modified model in its simulated test, with 50 samples per condition.Simulated harmful-request engagement50 samples per condition; engagement is not exploit success0%100%Direct request0%False cover story64%Prefilled reasoning92%Modified weights100%Source: Anthropic, 29 September 2026.
The chart shows responses to prompts in Anthropic’s simulation, not attack success on real networks.

Anthropic also tested a method it called abliteration, which alters the model to reduce refusal behaviour. Its team reported roughly 2,200 GPU hours and about $4,400 in compute cost for a full-model modification. Across three public harmful-request benchmarks, the altered model’s refusal rate fell sharply, but it did not fall to zero on every benchmark: Anthropic reported about 3%, 2% and 12% refusals. Tom’s Hardware noted a practical counterweight: running a large model at useful speed still requires substantial hardware or rental expense. Low marginal API cost for one smaller-model test is not the same as cheap, unlimited offensive capability.

The $20.40 experiment needs context

In a separate researcher-guided trial, Anthropic said GLM-5.3-Flash, the smaller version, chained exploits for two already known flaws, including a publicly disclosed Chrome vulnerability. The company reported about 20 minutes of human setup and eight hours of model work. At Z.ai’s listed API prices, it estimated a $20.40 model-use cost. This was an experimental price estimate, not a complete budget for finding, deploying or monetising a real intrusion.

The difference between known and unknown vulnerabilities also matters. Building a chain from public details of patched bugs is a different task from discovering a new weakness in a live product. Anthropic separately described a researcher-led sandbox trial in which the full model found previously unknown browser flaws and produced a proof of concept that accessed a test file. The company said it had disclosed the browser issues to the maintainer. The report does not establish a broad campaign of in-the-wild exploitation.

What the independent reports add

South China Morning Post reported the 50-versus-56 result and identified the gap between capability and safeguards as the concern. Tom’s Hardware examined the resource cost of running a modified large model, a check on any claim that the entire attack pipeline is inexpensive. The Next Web distinguished the simulated safeguard tests from the benchmark attempts and traced the earlier NIST assessment. These are original reports by separate publishers; their common source for the experimental figures remains Anthropic, whose role is stated here.

This attribution is especially important because Anthropic sells a competing AI system. Its tests are detailed and useful, but its commercial and policy interests are relevant context. NIST’s separate evaluation supplies an independent lens on capability. Meanwhile, the public reports do not show evidence that GLM-5.3 has already caused a particular breach. A responsible conclusion is that access to strong dual-use tools is widening while actual misuse and effective defences require continued observation.

What Indian technology teams should do

For Indian startups and enterprises, the practical issue is not the model’s nationality. It is the shrinking time between disclosure of a software flaw and the arrival of usable testing or exploit code. Security teams should maintain an inventory of exposed software, know which systems are internet-facing, rehearse patch prioritisation and validate fixes with controlled testing. These are general defensive steps, not evidence that any named Indian organisation has been attacked by GLM-5.3.

Procurement teams also need to distinguish an open-weight model used inside a company from a hosted model offered by a third party. If a vendor promises abuse monitoring, ask whether those controls run in the customer’s installation, in a central API, or both. A policy shown for one deployment might not apply to a copied set of weights. The related Lapaas Voice report on the AI safety accord explains why published evaluation methods and follow-through matter more than broad assurances.

For a separate example of why infrastructure boundaries matter, see Lapaas Voice’s reporting on the OpenAI training pause after a DNS escape. Its circumstances are different, but both cases show that safety claims must be tied to the actual deployed system and its operating environment. The earlier GLM-5.3 launch coverage provides product background; Anthropic’s September cyber tests are the new event here.

What remains uncertain

Several questions remain open. Anthropic’s report does not reveal the undisclosed browser flaws while maintainers work on them, so outsiders cannot yet evaluate those cases in full. It does not quantify the volume of actual malicious GLM-5.3 use. NIST’s four-month gap is an estimate derived from its particular benchmark mix, not a countdown to an attack or a promise that US models will retain an advantage. Neither source proves that every downloadable copy has the same guardrails after modification.

What can be said now is narrower and more useful. In Anthropic’s controlled experiments, GLM-5.3 performed close to restricted Mythos Preview on selected exploitation tasks, and prompt or weight changes undermined its refusal behaviour. NIST independently found a leading open-weight cyber model, though its broader score remained below the US frontier. The business consequence is a need for faster defensive testing and clearer model-access policies, while real-world harm claims must wait for evidence.

Frequently asked questions

Is GLM-5.3 as powerful as Claude Mythos Preview?

Anthropic found close scores on two selected cyber benchmarks. NIST’s wider evaluation still placed GLM-5.3 below the US frontier. Neither result establishes equality across every security task.

Did the model hack a real company?

The public Anthropic report describes sandboxed tests and researcher-guided exercises. It says a browser maintainer received disclosure of newly found flaws, but it does not document a confirmed criminal breach caused by GLM-5.3.

Does a 100% bypass figure mean every attack works?

No. The figure measures whether an altered model engaged with harmful prompts in one simulated 50-sample test condition. Engagement is not the same as a completed exploit, a successful intrusion or a universal bypass rate.

Why does open-weight distribution matter?

Downloadable weights can be copied and changed outside the original developer’s service. That can help legitimate research, while making central monitoring and recall less effective after release.

Sources and dates: Anthropic research, 29 September 2026; NIST CAISI, 17 September 2026; independent original reports from South China Morning Post, Tom’s Hardware and The Next Web linked above. Event date for this article: Anthropic’s 29 September publication.

Frequently Asked Questions

What did Anthropic find about GLM-5.3?

Anthropic’s Frontier Red Team reported 50 successful end-to-end ExploitBench attempts out of 410 for GLM-5.3, versus 56 for Claude Mythos Preview. On a separate 100-task binary exploitation sample, the scores were 4% and 6%.

Is GLM-5.3 the strongest open-weight cyber model?

An independent US government evaluation found GLM-5.3 was the strongest open-weight cyber model it had tested, but it still trailed the US frontier overall.

Has GLM-5.3 been used in a real cyberattack?

No. The findings come from controlled tests and do not establish that criminals used the model in an actual breach. Vendor results elsewhere need similar care, as with Gemini 4’s coding benchmarks.

Get the day’s top stories in your inbox

One concise email. No spam, unsubscribe anytime.