On October 28, 2026, Google DeepMind launched Gemini 4 Carbon, a model explicitly built for cyber defense: vulnerability detection, validation, and remediation. The same day, security firm Irregular published tests showing an open Chinese model, GLM-5.3, capable of running real cyberattacks autonomously, refusing none of them. For an SME, these two announcements say the same thing from two angles: AI-driven cybersecurity is scaling up, on both the attack and defense side.
At a glance
- Google DeepMind presented Gemini 4 Carbon on October 28, 2026 as its strongest model for defensive cybersecurity: vulnerability detection, controlled exploitation testing, and code remediation.
- Access is for now restricted to a small group of defenders through Google's "Fairwind Program"; there is no access yet through the Gemini app, the public API, or Gemini Enterprise.
- The same day, security firm Irregular published a report on GLM-5.3, an open model from Chinese lab Zhipu AI, able to complete 3 full cyberattack scenarios out of 13 tested on its own, refusing none of the offensive tasks it was given.
- The NIST (the US standards institute) called GLM-5.3 the most capable openly available cybersecurity model published so far, about four months behind the top closed US models.
- For an SME, the signal isn't "AI attacks" or "AI defends": both capabilities are advancing at the same pace, making vigilance over access and vendors an operational priority.
Gemini 4 Carbon: a model built for security teams
Gemini 4 Carbon follows Gemini 4 Argon, released a few weeks earlier. Google presents it as able to work autonomously for several days on complex engineering tasks, such as identifying and fixing the root cause of a bug in a large codebase.
On cybersecurity specifically, evaluated without guardrails in a controlled test setting, Google reports precise results: 88.2% success on SEC-bench Pro (vulnerability detection), 54.8% on ExploitGym (exploitation), and 71.0% on CWE-bench (fixing vulnerable code). On alignment, Google reports a 0.0% cheating rate on its internal live coding tests and roughly an 80% drop in high-severity misalignment alerts compared with its previous model.
Read with caution
These figures come from Google DeepMind's official blog post (blog.google) and from Artificial Analysis. Google itself notes its cybersecurity tests were run without guardrails, in a controlled setting, which does not reflect general public use. No fully independent body has yet reproduced these results exactly.
For now, access to Gemini 4 Carbon stays limited to a subset of selected defenders under Google's Fairwind Program. The company says it plans to gradually widen access to more trusted partners, with no set timeline for public availability. The model is not yet available through the Gemini app, the public API, or Gemini Enterprise.
GLM-5.3: the flip side, on the attack
On the same day, October 28, 2026, security firm Irregular published its test results on GLM-5.3, an open-weight model from Chinese lab Zhipu AI. In simulated networks, the model fully completed 3 out of 13 tested cyberattack scenarios, and refused none of the offensive tasks it was handed.
NIST called this result, as reported by Insurance Journal, the most capable openly available cybersecurity model published so far, about four months behind the top closed US systems. This isn't a coincidence of timing: open models are advancing fast, and once their weights are downloaded, their publisher can no longer control how they're used.
Closed model (Gemini 4 Carbon)
Access controlled by Google, limited to vetted defenders. The publisher can revoke access or patch the model afterward.
Open model (GLM-5.3)
Weights are freely downloadable. Once released, the publisher can no longer restrict use or pull the model from machines that already downloaded it.
This echoes a 2024 report from the US Commerce Department (NTIA) on open-weight models: releasing a model "cannot feasibly be reversed," and direct access to the weights "can allow individuals to strip" the safety measures built in by its publisher. The same report also acknowledges the value of open models for independent security research, which is exactly what let Irregular test GLM-5.3 without Zhipu AI's prior approval.
What this actually means for an SME
Most SMEs will never touch Gemini 4 Carbon or GLM-5.3 directly. But both announcements accelerate a trend that affects them directly: tools capable of finding flaws in a computer system are becoming more powerful and more accessible, in both directions.
Assume attackers are advancing too
AI-automated vulnerability scanning is no longer reserved for large companies. Security patches need to go out faster than before, across every exposed system (websites, APIs, VPNs, connected SaaS tools).
Check access rights and dormant accounts
An AI-equipped attacker looks first for the easiest path in: weak passwords, old accounts, shared access that was never revoked. A regular access audit remains the most cost-effective measure for an SME.
Track announcements from the AI vendors you use
If the company uses Gemini, Claude, or ChatGPT for sensitive tasks, it's worth regularly checking the access conditions and safety measures the vendor announces, since they keep changing with each release.
Don't mistake a defense tool for a turnkey solution
An AI model specialized in cybersecurity, closed or open, remains just one tool among others. It complements, but doesn't replace, basic security hygiene: backups, updates, two-factor authentication.
FAQ
Can an SME use Gemini 4 Carbon right now?
No, not directly. Access is currently limited to a small group of defenders selected through Google's Fairwind Program. The model is not available in the Gemini app, through the public API, or in Gemini Enterprise as of this article's publication date.
Does GLM-5.3 pose an immediate threat to SMEs?
The risk isn't that GLM-5.3 specifically targets one particular SME, but that tools capable of running automated cyberattacks are becoming more accessible across the whole ecosystem. That strengthens the case for basic measures: up-to-date security patches, strong passwords, and regularly tested backups.
Should companies stop using open-weight AI models as a precaution?
No. Open models also play a recognized role in security: they let independent researchers, like Irregular in this case, audit and test a model's real capabilities without needing its publisher's approval. The choice between an open and a closed model mainly depends on the intended use and the level of control wanted over data.
How does Gemini 4 Carbon differ from previous Gemini models?
Gemini 4 Carbon stands out for its unprecedented scores on specialized cybersecurity benchmarks (SEC-bench Pro, ExploitGym, CWE-bench) and for its claimed ability to work autonomously on long tasks, such as fixing a complex bug over several days. Previous Gemini versions weren't specifically positioned around cyber defense.
AI-driven cybersecurity is advancing on both sides of the fence, and that pace won't slow down. For an SME, the best response remains the same as before AI: solid basics, applied consistently. To go further on securing AI tools in your business, check out our LUWAI Mag resources or discover real-world cases in our customer stories.
