{"id":658513,"date":"2026-09-21T18:08:22","date_gmt":"2026-09-21T16:08:22","guid":{"rendered":"https://aivancity.ai/blog/?p=658513"},"modified":"2026-09-21T18:08:25","modified_gmt":"2026-09-21T16:08:25","slug":"glm-5-3-ce-que-ses-progres-en-codage-et-cybersecurite-permettent-reellement-daffirmer","status":"publish","type":"post","link":"https://aivancity.ai/blog/glm-5-3-ce-que-ses-progres-en-codage-et-cybersecurite-permettent-reellement-daffirmer/","title":{"rendered":"GLM-5.3 : ce que ses progrès en codage et cybersécurité permettent réellement d’affirmer"},"content":{"rendered":"\n
\n

\n Unveiled by Z.ai on \n August 14, 2026, GLM-5.3 retains the same \n base model as GLM-5.2 but has been enhanced through post-training \n focused on software engineering, \n long-term agentic tasks, and \n vulnerability research. Its published results are \n remarkable on certain tests, though they do not establish general \n or independent superiority.\n

\n\n

\n The real challenge lies as much in the dissemination of these capabilities as in their \n performance: how can we make an \n open-weight model useful to defenders without lowering the barriers for \n attackers?\n

\n
\n\n\n\n

What Just Happened

\n\n\n\n

Z.ai announced GLM-5.3 on August 14, 2026, as an evolution of GLM-5.2 designed for complex coding, tool use, and long-term agentic tasks. At the time of the announcement, the company had opted for a phased rollout and reserved the initial evaluation of the most sensitive cyber capabilities for selected partners. It indicated that it intended to spend approximately two weeks on evaluations and security hardening before a wider release.[1][3]

\n\n\n\n

This situation has since changed. As of September 21, 2026, Z.ai’s documentation indicates that GLM-5.3 is available to users of the GLM Coding Plan and that the model supports text input, a context window of one million tokens, and a maximum output of 128,000 tokens.[2] The weights are also available on Z.ai’s official Hugging Face account, under a license specific to GLM-5.3.[4] It is therefore no longer accurate to describe their release as a future milestone.

\n\n\n\n

The editorial significance of this announcement lies not only in a new code model. GLM-5.3 documents a broader shift: improvements to a model can now stem largely from post-training, learning environments, and agentic orchestration, without any announced changes to the base model. This advancement becomes particularly significant when it comes to the discovery and controlled exploitation of vulnerabilities.

\n\n\n\n

What's Really New

\n\n\n\n

According to Z.ai, GLM-5.3 uses the same base model as GLM-5.2. The claimed improvements stem from post-training—that is, the stages that follow general pre-training and specialize the system using examples, evaluation feedback, and interactive environments.[1][4] The innovation, therefore, does not lie in a completely redesigned architecture, but rather in a change in the capabilities achieved based on an existing foundation.

\n\n\n\n

This distinction is important. It suggests that a model’s size and the volume of its pre-training are no longer sufficient to describe its operational capability. For software engineering tasks, the quality of the environments, the duration of action trajectories, access to tools, and the reasoning budget can significantly alter the outcome. A model trained to navigate a repository, run tests, and correct errors is not evaluated as a simple generator of code snippets.

\n\n\n\n

The second new development is the reported progress on cybersecurity tasks that go beyond mere detection. Z.ai claims that performance improves as the evaluation approaches an exploitation chain. This claim must be viewed with caution: it is based primarily on the results and protocols published by the lab, rather than on an independent study covering the full range of professional use cases.[1][4]

\n\n\n\n\n\n
\n
\n\n
\n \"Executive\n
\n
\n\n
\n\n \"aivancity\"\n\n
\n

Executive MBA in AI & Business Transformation

\n

The MBA Redesigned for the Age of AI. For experienced executives who want to lead the transformation of their organizations. Paris, Nice, and Dubai.

\n
\n 12 months — Part-time\n At least 10 years of experience\n Early bird: 20,000 €\n Paris · Nice · Dubai\n
\n
\n\n
\n \n Learn more about the program →\n \n \n \n 100% in English\n \n
\n\n
\n
\n
\n\n\n\n

How GLM-5.3 Works

\n\n\n\n

GLM-5.3 is a large-scale, open-weight language model designed to reason and act in development environments. It receives an instruction, generates a sequence of steps, uses the tools at its disposal, and adjusts its actions based on the results obtained. In a supervised environment, it can, among other things, browse files, interact with a terminal, run tests, or examine a codebase.

\n\n\n\n

Z.ai presents the model as an evolution of GLM-5.2, with enhanced capabilities following pre-training. The official model published on Hugging Face is described as having approximately 753 billion parameters. The API documentation lists three levels of reasoning effort—low, high, and max—with reasoning remaining enabled.[2][4]

\n\n\n\n

In cybersecurity, the principle is not to provide the model with some mysterious knowledge of vulnerabilities. Rather, it involves exposing it to environments and tasks where it must analyze software, formulate hypotheses, verify whether a behavior is truly vulnerable, and, in certain controlled test environments, take validation a step further. The same mechanisms can support defensive auditing or, if misused, reduce the cost of certain offensive activities.

\n\n\n\n

Announced Features and Technical Limitations

\n\n\n\n
\n \n \n \n \n\n \n \n \n\n \n \n \n\n \n \n\n \n \n\n \n \n\n \n \n\n \n \n\n \n \n\n \n \n\n \n \n\n \n \n\n \n \n\n \n \n \n
\n Capacity or Constraint\n \n What You Need to Know\n
\n Agent-based modeling\n \n Analysis of deposits, use of tools, and performance of complex tasks,\n subject to the permissions granted.\n
\n Context\n \n Up to 1 million tokens, as stated in the Z.ai documentation,\n depending on the access method and terms of service.\n
\n Exit\n \n Up to 128,000 tokens have been announced, which can increase the duration,\n cost, and margin of error for long trajectories.\n
\n Reasoning\n \n Three performance levels are listed. Benchmark results often use\n the highest setting and high resource budgets.\n
\n Open weights\n \n Available on Hugging Face under a specific GLM-5.3 license,\n which is distinct from a standard open-source license.\n
\n Cybersecurity\n \n Research and validation of vulnerabilities in\n controlled environments, with a clear risk of dual use.\n
\n
\n\n\n\n

What Benchmarks Actually Tell Us

\n\n\n\n

The reported scores show a clear improvement for GLM-5.3 compared to GLM-5.2 under the conditions specified by Z.ai. On Terminal-Bench 3.0, the published score rose from 4.6 to 28.3. On DeepSWE v1.1, it rose from 46.2 to 66.9. On CyberGym, it reached 84.5, compared to 77.2 for GLM-5.2. ExploitBench improved from 24.4 to 54.4.[4]

\n\n\n\n
\n \n \n \n \n\n \n\n \n\n \n \n \n\n \n \n \n\n \n\n \n\n \n \n\n \n \n\n \n\n \n\n \n \n\n \n \n\n \n\n \n\n \n \n\n \n \n\n \n\n \n\n \n \n\n \n \n\n \n\n \n\n \n \n \n
\n Evaluation\n \n GLM-5.2\n \n GLM-5.3\n \n Read Carefully\n
\n Terminal-Bench 3.0\n \n 4,6\n \n 28,3\n \n Significant growth in the terminal environment, with a specific protocol and\n budget.\n
\n DeepSWE v1.1\n \n 46,2\n \n 66,9\n \n Improvements in software engineering, without any single model dominating\n the comparison.\n
\n CyberGym\n \n 77,2\n \n 84,5\n \n Highest score on the published chart for the discovery of\n vulnerabilities.\n
\n ExploitBench\n \n 24,4\n \n 54,4\n \n The score more than doubled, but it's lower than the 78.0 for Fable 5 and the\n 76.5 for GPT-5.6 Sol reported by Z.ai.\n
\n ExploitGym 2 hrs / 6 hrs\n \n 29 / 39\n \n 105 / 130\n \n Significant improvement, but still lagging behind the top proprietary models\n in this table.\n
\n
\n\n\n\n

These results support a limited conclusion: GLM-5.3 performs significantly better than GLM-5.2 on published tests and is on par with the best reported results for certain tasks. They do not support the conclusion that the model is generally better at coding or cybersecurity.

\n\n\n\n

Several precautions should be taken into account. Most of the results are published by Z.ai. The models are not always evaluated using identical budgets. Some tests allow for contexts of up to 400,000 or one million tokens, outputs of up to 128,000 tokens, multiple attempts, and processing times of up to several hours.[4] Cost , latency, reproducibility, the quality of corrections, and the false positive rate in production cannot be directly inferred from these scores.

\n\n\n\n

The 2,436 vulnerabilities are a warning sign, not a definitive confirmation

\n\n\n\n

Z.ai states that its work with Chinese cybersecurity teams has led to the identification of 2,436 vulnerabilities in 269 projects following human review, filtering, and deduplication. The lab has identified 107 critical vulnerabilities and 990 high-severity vulnerabilities. At the time of the announcement, only a portion of these had been made public, with the remainder remaining under embargo as part of a coordinated disclosure.[1][3]

\n\n\n\n

This volume is significant, but its evidential value depends on several pieces of information that are not fully available in the sources consulted: the project sampling method, the false-positive rate prior to review, the proportion attributable to GLM-5.3 rather than to earlier versions, reproducibility, corrections that were actually incorporated, and independent validation of each discovery. These figures should therefore be clearly attributed to Z.ai and not presented as an external measure of performance.

\n\n\n\n

The announced disclosure registry is nonetheless a useful practice if it allows each discovery to be linked to a project, a severity level, a fix status, and, where applicable, a CVE identifier. To assess its actual impact, it will be necessary to track the proportion of vulnerabilities confirmed by maintainers, the time to fix, and any independent publications.

\n\n\n\n

Why This Matters to Organizations

\n\n\n\n

For development and security teams, GLM-5.3 illustrates the shift from one-off assistance to carrying out longer-term tasks. An agent can theoretically scan a repository, formulate a hypothesis, trigger tests, and propose a fix. This continuity can increase audit coverage, particularly for open-source projects that lack resources.

\n\n\n\n

It also changes the expected skill set. Value no longer lies solely in writing code, but in defining permissions, designing isolated environments, verifying changes, and managing disclosure. Professionals must know how to supervise an agent, evaluate its evidence, and halt a trajectory when the system exceeds its authorized scope.

\n\n\n\n

Organizations should therefore not deploy this type of model with unrestricted access to the network, confidential information, or production environments. Responsible use requires sandboxes, least-privilege access, logging, human validation, and rollback procedures. The promised productivity gains do not absolve either the developer or the organization that authorizes the system’s actions of their responsibility.

\n\n\n\n

Issues of Accountability and Governance

\n\n\n\n

The first challenge is dual-use. The skills needed to find and validate a vulnerability can support defense, but they can also facilitate the search for an exploitation vector. The evaluations published by Z.ai show precisely that GLM-5.3 is not only making progress in detection. The distribution of weights also makes certain safeguards more difficult to enforce, since the model can be run and modified outside the provider’s infrastructure.

\n\n\n\n

The second issue concerns responsible disclosure. A discovery should not be made public until the maintainers have had a reasonable amount of time to understand and fix the problem. A system capable of generating numerous alerts requires appropriate governance: human triage, management of sensitive information, coordination with the affected projects, clear assignment of responsibilities, and the maintenance of verifiable records.

\n\n\n\n

The third issue is legal and contractual. Using an agent on a system without authorization may be unlawful, even if the stated purpose is research. Companies must also verify repository licenses, model terms of use, data processing, and the protection of trade secrets. In Europe, the applicable obligations depend on the actor’s role, the intended use, and the level of risk. The AI Act does not replace the GDPR, cybersecurity regulations, or contract law. This article does not constitute legal advice.

\n\n\n\n

Finally, the environmental cost must not be overlooked. Long-context evaluations, very long outputs, and agents that make multiple attempts consume more computational resources. The sources consulted do not provide sufficient data to quantify the specific environmental impact of GLM-5.3. This lack of measurement should be noted rather than replaced by a speculative estimate.

\n\n\n\n

What to Watch for Now

\n\n\n\n

The actual impact of GLM-5.3 will depend less on a single score than on four factors: the independent replication of results, the quality of confirmed vulnerabilities, the effectiveness of proposed fixes, and the governance of sensitive uses. It will also be necessary to distinguish between the technical availability of the weights and the practical accessibility of a model with approximately 753 billion parameters, the local operation of which requires significant infrastructure.[4]

\n\n\n\n

GLM-5.3 demonstrates that targeted post-training can significantly transform the capabilities of an existing model. It does not demonstrate that an open-weight model is, by nature, safer, more effective, or more useful than a closed service. Above all, it highlights an enduring challenge for agent-based AI: when defensive and offensive capabilities advance together, the quality of evaluation, permissions, and governance becomes just as critical as that of the model itself.

\n\n\n\n
\n\n

\n Learn more\n

\n\n

\n Check out our selection of articles to gain a better understanding of the evolution of open models, agentic coding, and AI-assisted cybersecurity:\n

\n\n \n
\n\n\n\n
\n\n

\n Sources\n

\n\n

\n [1] Z.ai. (2026).\n GLM-5.3: Frontier Coding with Emergent Cyber Capabilities.
\n \n https://z.ai/blog/glm-5.3\n \n

\n\n

\n [2] Z.ai Developer Documentation. (2026).\n GLM-5.3 Overview. Accessed September 21, 2026.
\n \n https://docs.z.ai/guides/llm/glm-5.3\n \n

\n\n

\n [3] Knight, W. (August 18, 2026).\n The Powerful Chinese AI Model Experts Warned About and Waited For Is Here.\n WIRED.
\n \n https://www.wired.com/story/zai-open-weight-ai-models-release-cybersecurity-hacking/\n \n

\n\n

\n [4] Z.ai. (2026).\n GLM-5.3, official model documentation on Hugging Face.\n Accessed September 21, 2026.
\n \n https://huggingface.co/zai-org/GLM-5.3\n \n

\n\n

\n [5] Axios. (August 14, 2026).\n A Chinese Lab’s New Model Is Nearly as Good at Hacking as U.S. AI.
\n \n https://www.axios.com/2026/08/14/china-open-source-ai-glm-53\n \n

\n\n
\n","protected":false},"excerpt":{"rendered":"

Unveiled by Z.ai on August 14, 2026, GLM-5.3 retains the same base model as GLM-5.2 but has been enhanced through post-training focused on software engineering, long-term agentic tasks, and research into…

\n","protected":false},"author":7,"featured_media":658515,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"content-type":"","footnotes":""},"categories":[63],"tags":[59],"class_list":["post-658513","post","type-post","status-publish","format-standard","has-post-thumbnail","category-ia-agentique","tag-parlonsia"],"yoast_head":"\nGLM-5.3: What Its Advances in Coding and Cybersecurity Actually Tell Us\n\n\n\n\n\n\n\n\n\n\n\n\n\t\n\t\n\t\n\n\n\n\t\n\t\n\t\n\n","yoast_head_json":{"title":"GLM-5.3 : ce que ses progrès en codage et cybersécurité permettent réellement d’affirmer","description":"GLM-5.3 is making progress in coding and cybersecurity through post-training. Analysis of benchmarks, limitations, applications, and dual-use risks.","robots":{"index":"index","follow":"follow","max-snippet":"max-snippet:-1","max-image-preview":"max-image-preview:large","max-video-preview":"max-video-preview:-1"},"canonical":"https://aivancity.ai/blog/glm-5-3-ce-que-ses-progres-en-codage-et-cybersecurite-permettent-reellement-daffirmer/","og_locale":"fr_FR","og_type":"article","og_title":"GLM-5.3 : ce que ses progrès en codage et cybersécurité permettent réellement d’affirmer","og_description":"GLM-5.3 progresse en codage et cybersécurité grâce au post-entraînement. Analyse des benchmarks, limites, usages et risques de double usage.","og_url":"https://aivancity.ai/blog/glm-5-3-ce-que-ses-progres-en-codage-et-cybersecurite-permettent-reellement-daffirmer/","og_site_name":"aivancity blog","article_published_time":"2026-09-21T16:08:22+00:00","article_modified_time":"2026-09-21T16:08:25+00:00","og_image":[{"width":1024,"height":1024,"url":"https://aivancity.ai/en/blog/wp-content/uploads/2026/09/lKimi-K3Architecte-Cloud1.png","type":"image/png"}],"author":"aivancity","twitter_card":"summary_large_image","twitter_misc":{"Écrit par":"aivancity","Durée de lecture estimée":"12 minutes"},"schema":{"@context":"https://schema.org","@graph":[{"@type":"Article","@id":"https://aivancity.ai/blog/glm-5-3-ce-que-ses-progres-en-codage-et-cybersecurite-permettent-reellement-daffirmer/#article","isPartOf":{"@id":"https://aivancity.ai/blog/glm-5-3-ce-que-ses-progres-en-codage-et-cybersecurite-permettent-reellement-daffirmer/"},"author":{"name":"aivancity","@id":"https://aivancity.ai/blog/#/schema/person/70f8508e84e45571c5fd172ea40ef3d4"},"headline":"GLM-5.3 : ce que ses progrès en codage et cybersécurité permettent réellement d’affirmer","datePublished":"2026-09-21T16:08:22+00:00","dateModified":"2026-09-21T16:08:25+00:00","mainEntityOfPage":{"@id":"https://aivancity.ai/blog/glm-5-3-ce-que-ses-progres-en-codage-et-cybersecurite-permettent-reellement-daffirmer/"},"wordCount":2465,"image":{"@id":"https://aivancity.ai/blog/glm-5-3-ce-que-ses-progres-en-codage-et-cybersecurite-permettent-reellement-daffirmer/#primaryimage"},"thumbnailUrl":"https://aivancity.ai/blog/wp-content/uploads/2026/09/lKimi-K3Architecte-Cloud1.png","keywords":["Parlons IA"],"articleSection":["IA Agentique"],"inLanguage":"fr-FR"},{"@type":"WebPage","@id":"https://aivancity.ai/blog/glm-5-3-ce-que-ses-progres-en-codage-et-cybersecurite-permettent-reellement-daffirmer/","url":"https://aivancity.ai/en/blog/glm-5-3-ce-que-ses-progres-en-codage-et-cybersecurite-permettent-reellement-daffirmer/","name":"GLM-5.3 : ce que ses progrès en codage et cybersécurité permettent réellement d’affirmer","isPartOf":{"@id":"https://aivancity.ai/blog/#website"},"primaryImageOfPage":{"@id":"https://aivancity.ai/blog/glm-5-3-ce-que-ses-progres-en-codage-et-cybersecurite-permettent-reellement-daffirmer/#primaryimage"},"image":{"@id":"https://aivancity.ai/blog/glm-5-3-ce-que-ses-progres-en-codage-et-cybersecurite-permettent-reellement-daffirmer/#primaryimage"},"thumbnailUrl":"https://aivancity.ai/blog/wp-content/uploads/2026/09/lKimi-K3Architecte-Cloud1.png","datePublished":"2026-09-21T16:08:22+00:00","dateModified":"2026-09-21T16:08:25+00:00","author":{"@id":"https://aivancity.ai/blog/#/schema/person/70f8508e84e45571c5fd172ea40ef3d4"},"description":"GLM-5.3 is making progress in coding and cybersecurity through post-training. Analysis of benchmarks, limitations, applications, and dual-use risks.","breadcrumb":{"@id":"https://aivancity.ai/blog/glm-5-3-ce-que-ses-progres-en-codage-et-cybersecurite-permettent-reellement-daffirmer/#breadcrumb"},"inLanguage":"fr-FR","potentialAction":[{"@type":"ReadAction","target":["https://aivancity.ai/blog/glm-5-3-ce-que-ses-progres-en-codage-et-cybersecurite-permettent-reellement-daffirmer/"]}]},{"@type":"ImageObject","inLanguage":"fr-FR","@id":"https://aivancity.ai/blog/glm-5-3-ce-que-ses-progres-en-codage-et-cybersecurite-permettent-reellement-daffirmer/#primaryimage","url":"https://aivancity.ai/en/blog/wp-content/uploads/2026/09/lKimi-K3Architecte-Cloud1.png","contentUrl":"https://aivancity.ai/blog/wp-content/uploads/2026/09/lKimi-K3Architecte-Cloud1.png","width":1024,"height":1024},{"@type":"BreadcrumbList","@id":"https://aivancity.ai/blog/glm-5-3-ce-que-ses-progres-en-codage-et-cybersecurite-permettent-reellement-daffirmer/#breadcrumb","itemListElement":[{"@type":"ListItem","position":1,"name":"Accueil","item":"https://aivancity.ai/blog/"},{"@type":"ListItem","position":2,"name":"GLM-5.3 : ce que ses progrès en codage et cybersécurité permettent réellement d’affirmer"}]},{"@type":"WebSite","@id":"https://aivancity.ai/blog/#website","url":"https://aivancity.ai/en/blog/","name":"aivancity blog","description":"Advancing Education in Artificial Intelligence","potentialAction":[{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https://aivancity.ai/blog/?s={search_term_string}"},"query-input":{"@type":"PropertyValueSpecification","valueRequired":true,"valueName":"search_term_string"}}],"inLanguage":"fr-FR"},{"@type":"Person","@id":"https://aivancity.ai/blog/#/schema/person/70f8508e84e45571c5fd172ea40ef3d4","name":"aivancity","image":{"@type":"ImageObject","inLanguage":"fr-FR","@id":"https://secure.gravatar.com/avatar/0b60f844cf48367ece3a9988562f25406b914c56b83ccd3df68e4c07737dc27e?s=96&d=mm&r=g","url":"https://secure.gravatar.com/avatar/0b60f844cf48367ece3a9988562f25406b914c56b83ccd3df68e4c07737dc27e?s=96&d=mm&r=g","contentUrl":"https://secure.gravatar.com/avatar/0b60f844cf48367ece3a9988562f25406b914c56b83ccd3df68e4c07737dc27e?s=96&d=mm&r=g","caption":"aivancity"},"url":"https://aivancity.ai/en/blog/author/bouazizaivancity-ai/"}]}},"amp_enabled":true,"_links":{"self":[{"href":"https://aivancity.ai/blog/wp-json/wp/v2/posts/658513","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https://aivancity.ai/blog/wp-json/wp/v2/posts"}],"about":[{"href":"https://aivancity.ai/blog/wp-json/wp/v2/types/post"}],"author":[{"embeddable":true,"href":"https://aivancity.ai/blog/wp-json/wp/v2/users/7"}],"replies":[{"embeddable":true,"href":"https://aivancity.ai/blog/wp-json/wp/v2/comments?post=658513"}],"version-history":[{"count":3,"href":"https://aivancity.ai/blog/wp-json/wp/v2/posts/658513/revisions"}],"predecessor-version":[{"id":658517,"href":"https://aivancity.ai/blog/wp-json/wp/v2/posts/658513/revisions/658517"}],"wp:featuredmedia":[{"embeddable":true,"href":"https://aivancity.ai/blog/wp-json/wp/v2/media/658515"}],"wp:attachment":[{"href":"https://aivancity.ai/blog/wp-json/wp/v2/media?parent=658513"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https://aivancity.ai/blog/wp-json/wp/v2/categories?post=658513"},{"taxonomy":"post_tag","embeddable":true,"href":"https://aivancity.ai/blog/wp-json/wp/v2/tags?post=658513"}],"curies":[{"name":"wp","href":"https://api.w.org/{rel}","templated":true}]}}