Chinese AI lab says its new coding model got scarily good at finding security flaws, almost by accident

[StoneCold_99]

Chinese AI developer Zhipu has launched GLM-5.3, a coding focused model the company says developed unexpectedly strong cybersecurity capabilities as a side effect of training it to be a better programmer. On CyberGym, a benchmark testing vulnerability identification, GLM-5.3 scored 84.5%, edging out Anthropic's Mythos 5 at 83.8% and OpenAI's GPT-5.6 Sol at 83.6%. It trails both by a much wider margin on ExploitBench, which tests actually exploiting a found vulnerability, scoring 54.4% against 78% and 76.5% respectively.

Zhipu says the model moved beyond spotting isolated bugs to forming coherent plans for complete exploitation chains, and attributes the jump specifically to scaling post training, reinforcement learning across increasingly complex simulated work environments, rather than building an entirely new base model. The company's ExploitBench score more than doubled from GLM-5.2's 24.4%, and on a separate benchmark testing sustained task completion, GLM-5.3 finished 105 exploitation tasks within two hours versus 29 for its predecessor.

Working with security teams in China, Zhipu ran the model against real world codebases and says it identified 2,436 vulnerabilities across 269 projects after expert review and deduplication, including 1,097 medium to high severity issues spanning system kernels, browser engines, and network protocols. The oldest flaw found dated back to 1981, and vulnerabilities in the dataset had sat undiscovered in code for an average of 26.6 years.

Neil Shah of Counterpoint Research frames the underlying issue plainly, saying the same reasoning an AI uses to test code and fix bugs is exactly what an attacker uses to find a weak spot and break through it, meaning offensive cyber capability is becoming an inherent byproduct of training a genuinely good coding model rather than something that has to be deliberately built in. That's a real problem specifically because Zhipu plans to release GLM-5.3's model weights openly about two weeks after launch, meaning any safety guardrails built in during training could potentially be stripped out once the weights are public.

So the honest tension here is real. The same capability that lets the model find and help fix thousands of genuine security flaws in existing software is functionally identical to the capability that could let a bad actor find and exploit them, and once the weights are open, that capability is available to absolutely anyone regardless of intent

Zhipu says new coding AI developed advanced cyber skills faster than expectedwww.infoworld.com

[Fox50]

Zhipu scoring ahead on vulnerability discovery but well behind on actual exploitation is a clearly important distinction that headlines are going to flatten.

Finding a weak spot and actually breaking through it are meaningfully different skills, and the gap between them still matters a lot for real world risk

[Freddie]

Wondering how Western labs are actually handling this exact same tension internally.

If training for coding competence reliably produces offensive security capability as a side effect, that's presumably showing up in Anthropic and OpenAI's own internal testing too. That part surprised me

[Cameron83]

The open weight release timeline is the part that should worry people most. Two weeks between launch and public weights is not a long window for safety hardening given what Shah is describing as an inherent, hard to strip capability

[Henry25]

Well reported! for anyone in security or software engineering to actually read in full, the framing of the response window shrinking to near zero once this capability is freely downloadable is a really sobering way to think about what open weight releases mean going forward

[HeartbreakKidCurtis18]

Extra point, the responsible disclosure process Zhipu describes.

Expert review before public release, embargo on most findings, is at least the right instinct even if the underlying dual use problem remains completely unsolved once the weights go open

[Puma22]

Wonder what actual technical safeguards Zhipu has planned beyond the vague safety evaluation and hardening mentioned.

That phrase is doing a lot of work in the announcement without much concrete detail behind it

[NovaPrime68]

The reinforcement learning across complex task environments explanation for the improvement is the more technically interesting part for anyone following how these models actually get trained. This isn't really about a bigger model, it's about a better training process

[Jackson1]

This tracks with a broader pattern a quite capable coding model and a particularly capable hacking tool are turning out to be much closer to the same thing than most people assumed when these coding assistants first started getting popular. Small but real thing

[Maisie84]

2, 436 vulnerabilities across 269 projects found by a single model run is a enormous number to process responsibly.

Makes you wonder whether the human security teams reviewing these findings can actually keep pace with the rate AI can generate them

[Cyborg77]

The 1981 vulnerability and 26.6 year average undiscovered age is the most striking detail in the whole story. That's not a hypothetical risk, that's real long buried exposure sitting in production systems for literal decades