Cisco open-sources tiny Antares models to locate code flaws
Cisco's Foundation AI arm releases 350M and 1B open-weight models that trace known vulnerabilities to specific files, nearing frontier accuracy at a fraction of the cost.
Cisco's Foundation AI team has released Antares, a family of small open-weight language models built for one job: vulnerability localization — connecting public advisories, CVE records, and Common Weakness Enumeration entries to the specific files in a repository where the flaw likely lives. The 350M- and 1B-parameter versions are now available on Hugging Face.
An investigator, not a scanner
Rather than pattern-matching over code, each Antares model works a repository the way a human security analyst would: it starts from a vulnerability description, searches for matching code patterns, reads candidate files, weighs new evidence, and backtracks when a lead runs dry. To evaluate the approach, Cisco built a 500-entry Vulnerability Localization Benchmark that forces a model to navigate unfamiliar codebases and recognize weakness patterns by category. On that benchmark, Cisco says the 1B model approaches the accuracy of frontier systems like GPT-5.5 while remaining small enough to run cheaply on local hardware — a meaningful property for security teams that cannot ship proprietary source code to external APIs.
The release extends Cisco's Foundation AI open-model program into a week thick with AI-security news, and the company is explicit about the economics: triaging a steady stream of vulnerability reports with a frontier model is expensive at scale, while a specialized 1B model makes continuous, repository-wide localization affordable.
Why it matters
Antares is a data point in two live debates. First, whether small specialized models can beat frontier generalists on narrow high-value tasks — Cisco's answer is yes, by orders of magnitude on cost. Second, how defenders keep pace as AI systems find vulnerabilities faster than humans can fix them, a concern Google raised the same week with its own security-model push. Open weights matter here: security tooling that runs locally, audits transparently, and costs little to operate is far easier for enterprises and governments to adopt than a black-box API — and it hands the defensive side of the AI-security arms race a genuinely commodity-priced weapon.