“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.
Antares-350M and Antares-1B: small models, narrow mission
Cisco has released two small language models built specifically to hunt known vulnerabilities in codebases: Antares-350M and Antares-1B. The models are part of a new Antares family of security small language models (SLMs) and are available on Hugging Face — but access is restricted to vetted users. Cisco said it is working with academic and nonprofit organizations, as well as smaller and public organizations’ security teams, to grant access in a controlled way.
Local execution, source-code keys, and a privacy argument
Cisco framed Antares as a tool designed to run locally inside organizations. “Because both are small models designed to run locally, ‘you also need the keys to the source code’ to scan for and find vulnerabilities,” Sampath said. The company contrasted this with cloud-based LLMs that send code to external servers, arguing that local operation keeps proprietary code on an organization’s machines and enables analysis in environments with strict privacy or compliance requirements. Sampath warned that requiring source-code keys is also a practical barrier: “This means an attacker is going to be able to exploit an endpoint or a service that you have,” underscoring that local access to code remains a prerequisite for the scanner to work.
Benchmark claims and cost-performance comparisons
Cisco claims Antares models perform as well as or better than dozens of larger models on a new benchmark the company developed to measure how efficiently AI models identify security flaws in codebases. According to Cisco’s public statements, Antares-1B outperforms Google’s Gemini 3 Pro and is comparable to Z.ai’s GLM-5.2. The company also said a not-yet-public Antares-3B outperforms GLM-5.2 and OpenAI’s GPT-5.5 in finding vulnerabilities.
Cisco provided two concrete operational comparisons: speed and estimated cost. Amin Karbasi, Cisco vice president and chief AI scientist, said Antares can finish scanning a cohort of 500 repositories in 15 minutes, whereas “frontier models take five hours.” Karbasi added that this speed translates to much lower compute costs: “It takes like less than $1 whereas frontier models are above $100 into $150 of cost.”
Design philosophy: investigator, not chatbot
Cisco emphasizes a different training approach and a focused design. “Antares is inherently not a chatbot. It is an investigator. It is a search engine,” Karbasi told The Register. He said the models were trained to use multiple search strategies because “one way of search may not actually be fruitful, then it has to change its strategy, do it another way, and then do it another way.” The company argues that the smaller model size makes Antares “nimble” and able to run many searches in parallel, a contrast Karbasi illustrated with two analogies: a bicycle on a busy London street that can “go much faster than the biggest truck,” and the name Antares — after the red super-giant star — to emphasize that a tiny part of code can have outsized impact: “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.”
What this means for technologists, security teams, and adversaries
- Technologists and security teams: Antares offers a locally runnable scanner that Cisco says identifies vulnerabilities quickly and cheaply. Organizations with strict privacy or compliance constraints will likely watch vetted availability and local deployment guarantees closely.
- Procurement and compliance officers: The gated-release model and Cisco’s stated partnerships with academic and nonprofit groups will shape who can access Antares models and under what terms — particularly relevant for organizations weighing cloud-based versus on-premises code analysis.
- Adversaries and threat actors: Cisco’s point that the scanner requires access to the source code serves as a reminder that misuse depends on having local or privileged access; Cisco’s gating and local-execution design aim to limit the tool’s availability and reduce cloud-exfiltration risk.
Cisco is positioning Antares as a new class of compact, purpose-built security models: limited in scope but optimized for scanning speed, cost, and local privacy. The company is restricting wider distribution — notably keeping a planned 3-billion-parameter Antares-3B model off public release and “completely gating” it, Karbasi said, to ensure it is “responsibly release[d] to communities that need it.” Whether that gated, locally run approach will reconcile rapid vulnerability discovery with controlled access is the practical question the company has placed front and center.
