Cisco just dropped two small AI models built to sniff out software bugs. The networking giant named them Antares after a giant star. And they pack a surprising punch against far larger systems from Google and OpenAI.
The models, called Antares-350M and Antares-1B, belong to a new family of security-focused small language models. Cisco made them open-weight. Anyone vetted can grab them from Hugging Face right now. A bigger 3-billion-parameter version stays locked down for now. The company wants to hand it only to groups that will handle it carefully.
“We’re making sure we’re gating that and appropriately granting access,” DJ Sampath, Cisco’s senior vice president and general manager of AI software and platform, told The Register. Sampath stressed the models run locally. That keeps proprietary code inside the organization. No need to ship it off to some distant cloud server.
Attackers still need keys to the source code to make these models useful against a target. Without them, the AI can’t scan effectively. This setup appeals to universities, nonprofits, public sector teams and smaller outfits with tight compliance rules. Privacy stays intact. Costs stay low.
Amin Karbasi, Cisco VP and chief AI scientist, offered a vivid comparison. The star Antares measures nearly 1,000 times bigger than the sun yet appears small in the sky. Vulnerabilities work the same way. One flaw buried in a million lines of code can cause massive damage. “The impact of vulnerabilities in your codebase is huge, but it might be only a single file or a few lines of code in a million lines of code,” he told The Register.
Karbasi explained the models don’t act like chatbots. They function as investigators. Search engines. They hunt a specific needle in a haystack. Training involved multiple search strategies. One approach fails. The model shifts tactics. It repeats until it succeeds. Small size helps here. The system runs many searches at once. Bigger models lumber along with single massive passes.
Performance numbers back the approach. Cisco built a new benchmark with 500 tasks focused on vulnerability localization. The test measures how well AI connects external bug descriptions to specific files inside large codebases. Existing tests like SWE-Bench or CodeScout miss this angle. They emphasize other coding challenges.
In Cisco’s tests, Antares-1B beats Google’s Gemini 3 Pro. It matches Z.ai’s GLM-5.2. The unreleased Antares-3B tops both GLM-5.2 and OpenAI’s GPT-5.5 at finding flaws. These small models finish scans quickly. They analyzed 500 repositories in 15 minutes. Frontier models needed five hours. Cost tells an even sharper story. Antares ran the job for less than a dollar. Larger systems ran up bills between $100 and $150.
“If you look at the performance, in terms of the time it takes to finish 500 repositories, Antares finishes the entire cohort of repositories in 15 minutes, whereas frontier models take five hours,” Karbasi said in The Register. He added the cost difference makes continuous scanning realistic for more teams.
The blog post announcing the work includes praise from academics. Reza Shokri from the National University of Singapore noted that software always ships with bugs. Legacy code carries hidden risks. AI now writes more code and finds ways to attack it. Small models that run locally and check output in real time become essential. “Antares-1b is a good example: it recognizes vulnerabilities in code it is shown, across several weakness classes and languages,” Shokri said according to Cisco’s blog.
Amin Saberi from Stanford went further. Advanced detection shouldn’t remain a luxury for rich organizations. “Antares’s results change that equation: near-frontier accuracy on secure code reasoning at a fraction of the cost, fast enough to run on every commit,” he stated in the same Cisco announcement. At a time when attackers already use AI, efficiency turns into a requirement.
Cisco designed Antares around an iterative search process. The model starts with a vulnerability description. It searches for patterns. It reads files. It gathers evidence. When one path stalls, it pivots. This back-and-forth mimics how human analysts work. The output ranks likely files and provides a trace of its reasoning. Teams can use it to locate flaws by Common Weakness Enumeration type. They can triage security advisories. They can strengthen static analysis tools. They can add checks inside CI/CD pipelines.
But. The models only detect known bug patterns. They don’t invent new zero-days. They connect public vulnerability knowledge to private code. That distinction matters. Security teams still need human oversight. They need other tools. Yet the speed and price open doors for constant monitoring that bigger systems make impractical.
Karbasi compared the setup to riding a bicycle through busy London traffic. The bike slips ahead of the largest truck. Sampath offered his own image. “Sometimes you don’t need a private jet to go to a corner store, right?” he asked in The Register.
These releases fit a wider Cisco push. The company created the Foundry Security Spec, a blueprint for testing agentic AI systems on security tasks. It also offers CodeGuard, a set of rules for writing safer code. Antares adds concrete models and that new benchmark to the mix. The technical paper detailing the benchmark sits on Cisco’s site for those who want to dig deeper.
Industry watchers took notice immediately. On the day of release, posts on X echoed the original reporting from The Register. No major new studies appeared in the past 24 hours. Yet the conversation around open-weight security models keeps building. Chinese models have shown competitive results lately. Smaller specialized systems challenge the assumption that only the biggest players can deliver useful AI for defense.
Questions remain about long-term maintenance. How often will Cisco update the models as new weakness types emerge? Will the community contribute improvements since the weights are public? The gating on the 3B version suggests caution. Cisco learned from past AI releases where models were misused.
Still, the bet looks clear. Organizations don’t need frontier-scale compute for every security job. A focused, nimble model that runs on premises can handle the bulk of routine vulnerability hunting. It keeps data private. It slashes expenses. It runs fast enough to scan before code merges.
Security teams at smaller outfits may soon run Antares alongside their existing scanners. Larger enterprises might test it in air-gapped environments. Researchers could build on the benchmark to push the field forward. The star analogy holds. Sometimes the biggest impact comes from something that doesn’t dominate the sky.