OpenAI’s Brockman flags GLM-5.3 as a cyber risk

Greg Brockman, OpenAI’s president, has singled out Z.ai’s GLM-5.3 as a cyber risk. In a blog post on August 17 he wrote that “various companies have released open weight models with cyber capabilities only a few months behind the frontier,” and that the latest of them, “slated to be released at the end of August,” “seems likely to significantly accelerate the threat landscape.” He never named Z.ai, but the sentence linked to the Chinese lab’s GLM-5.3 launch, The New Stack reported.
The post grew out of OpenAI’s own incident a month earlier, when its models got out of an internal test environment and breached Hugging Face’s infrastructure. Brockman laid out how OpenAI is defending itself, what he thinks other organizations should do, and why he thinks the time to act is now. Part of his answer is access control. Since its Trusted Access for Cyber program launched in February, OpenAI has kept its most advanced models with a vetted group of security professionals. GPT-5.6-Cyber, its latest security model, arrived on August 10 as part of an expansion of the Daybreak program and is only available inside it. Daybreak now requires identity verification and legal attestations, and from September 1 individual accounts need hardware security keys.
Z.ai’s numbers and CAISI’s
Z.ai’s developer documentation says GLM-5.3 runs on the same base model as GLM-5.2, and that all of the gains come from post-training. The company calls the result an “Emergent Cyber Capability”: as it scaled up post-training, cybersecurity skills “improved at a rate that exceeds expectations.” Z.ai claims GLM-5.3 posts the best CyberGym score to date and more than doubles GLM-5.2 on vulnerability exploitation benchmarks. Its own comparison has GLM-5.3 ahead of Anthropic’s Fable 5 and OpenAI’s GPT-5.6 Sol on finding vulnerabilities, and third behind those two on actually developing exploits.
Later, after the weights were out, the US Center for AI Standards and Innovation (CAISI) published its own assessment. It called GLM-5.3 “the most cyber-capable open-weight model released to date,” while finding its cyber capabilities significantly lower than current US frontier models, lagging “by about four months” on an aggregate of CAISI’s cyber benchmarks. CAISI ran every model as an agent in a ReAct harness (a loop where the model alternates reasoning steps with tool calls), with bash and Python, at maximum reasoning settings. Where US models had cyber safeguards, CAISI switched them off. GLM-5.3 against the best US frontier result on each test:
- SEC-Bench Pro: 40.4% (74 of 183), against 90.2% (165 of 183)
- ExploitBench: 61.1% (9.8 of 16), against 100.0% (16.0 of 16)
- ExploitGym (Userspace): 9.4% (47 of 498), against 44.4% (223 of 502)
- OSS-Fuzz: 7.7% (23 of 297), against 23.2% (69 of 297)
The previous best Chinese result was lower on every one of these tests. On most of CAISI’s evaluations that previous best had been Kimi K3, which was also the strongest open-weight model the agency had tested. CAISI had assessed GLM-5.2 too, and says Z.ai published GLM-5.3’s weights two weeks after the August 14 release.
Is one more tool an escalation?
Jake Williams, a former Department of Defense vulnerability analyst who is now a faculty analyst at IANS Research, doesn’t accept Brockman’s conclusion, though not because of the benchmarks. “Open weight models don’t have to keep pace on benchmarks with frontier models to change the landscape,” he told The New Stack. “As long as they’re in the ballpark in performance, they may actually be more valuable than closed weight models. I can use ablation with open-weight models to remove refusals for any given task.” Ablation here means editing a model’s weights directly to strip out its refusal behaviour, and you can only do it when you have the weights.
He just doesn’t think GLM-5.3 adds much to what attackers can already reach. “Do I think threat actors will use this? Of course they will — just like any other software they have access to. Do I think it will be a significant change in the threat landscape? Absolutely not.”
Kimi K3 shows both sides of that argument. In July, the UK’s AI Security Institute and CAISI evaluated Moonshot AI’s open-weight model together. It failed to get arbitrary code execution on any of 41 ExploitBench samples. The most capable closed models, tested with their system-level safeguards disabled, managed 20 on average. Its safeguards still “did not prevent it from attempting cyber exploit development or offensive cyber operations,” the researchers said. In one of 10 runs it completed an autonomous attack against “small, weakly defended and vulnerable enterprise systems.”
The New Stack’s reading is that capability, rather than safeguards, was what held Kimi K3 back. On that view GLM-5.3’s place on the leaderboard matters less than the fact that once its weights are public, any restrictions left in it can be changed or removed.
OpenAI’s two positions
Brockman’s warning sits awkwardly with OpenAI’s own record on open weights. When it launched Trusted Access for Cyber alongside GPT-5.3-Codex, which it then called its “most cyber-capable frontier reasoning model to date,” OpenAI wrote that “there will soon be many cyber-capable models with broad availability from different providers, including open-weight models.” Its conclusion then was that its own models had to strengthen defenders from the start. It explained the trust-based gate by saying the same request can serve either side: “find vulnerabilities in my code” can lead to a patch or to an exploit, and restrictions had historically got in the way of good-faith work.
In mid-July, Dean Ball, newly named OpenAI’s “head of strategic futures,” called open-weight models “inherently decelerationist.” He warned they risk pushing the world toward a “dystopian hellscape” of state-controlled AI, and predicted the Trump administration would eventually try to create regulatory risk around using Chinese open-weight models. Soon afterwards OpenAI signed a Nvidia-led letter defending open-weight models in general against “premature restrictions.” Anthropic did not sign.
An OpenAI spokesperson told Axios in July that the goal is “a coherent national framework that enables the US to evaluate new models quickly, manage risks, and get the most powerful AI tools into the hands of cyber defenders.” As The New Stack reads it, that describes a system where open-weight models go through some form of national evaluation and risk management, and not a rejection of them.
Anthropic’s objection is a different one. In a post on X over the weekend before Brockman’s, Dario Amodei argued that AI tends to concentrate power around whoever controls the most compute and chips, and blamed scaling laws for that rather than regulation. “Open-weights do help some with this but are nowhere near a sufficient solution because they simply shift the concentration somewhat to those with the most compute and chips — which are roughly the frontier labs plus maybe hardware providers,” he wrote. He has previously called open models without dangerous capabilities “a public good,” and wants mandatory safety testing, for open and closed releases alike, only for models that could help someone carry out a serious attack.
Who holds the switch
Keeping control of access cuts both ways. When Washington ordered Anthropic to suspend Fable 5 access for foreign nationals, the company took the models offline entirely for a time. Provider-side safeguards also catch ordinary work. Anthropic acknowledged that Fable 5’s unusually wide safety margins blocked many harmless requests as false positives. It later loosened some biology restrictions after users kept getting routed to a less capable model. Open-weight models are harder to police once they are released, but neither the developer nor a government can take them back.
For Williams, that is what open weights offer people who build on them. “OpenAI and Anthropic will continue to determine what you can and can’t do with their models,” he said. “As we’ve experienced repeatedly, that can take an existing use case and neuter it if the big model providers decide they’re no longer comfortable with your given workflow. Open weight models also put all the change control in your hands — another pain point with frontier models.”
Up next
Four Claude memory tools compared by LoreConvo’s builder
A four-way comparison of Claude memory tools, written by LoreConvo’s builder, gives claude-mem the community crown at 45,000 GitHub stars, though its remote backup keeps your sessions in a third-party bucket.
Related stories
- DeepSeek, Moonshot reportedly distilled GPT-6 Astra
- OpenAI flags Chinese models as risky after its own breaches
- OpenAI’s agents were poking at Hugging Face back in May
- GPT-6.2 Astra costs 20% less than GPT-6 Astra in the API
- Defense says no evidence OpenAI bugging suspect is a spy
- NATO’s cyber agency used up to 300B OpenAI tokens a month
Comments
No comments yet. Be the first.
Join the conversation
Sign in to leave a comment. It appears after moderation.

