Cisco’s Antares AI: Revolutionizing Vulnerability Localization in Code

Bridging the Gap: AI for Software Vulnerability Detection

For readers tracking the shift, In the complex world of software development, identifying and patching vulnerabilities is a constant race. Developers and security teams often face the daunting task of sifting through vast, unfamiliar codebases to pinpoint the exact location of a known flaw, a process that is both time-consuming and expensive. Cisco Foundation AI is stepping up to this challenge with the release of Antares, a new family of security-focused Small Language Models (SLMs) designed specifically for vulnerability localization.

Meanwhile, Antares aims to streamline the critical first triage step in software security: given a description of a vulnerability and a repository, it efficiently identifies the files containing the flaw. This targeted approach promises to significantly reduce the manual effort involved in connecting external security intelligence with internal source code.

Introducing Antares: Targeted Security SLMs

Cisco Foundation AI has made two key Antares models, Antares-350M and Antares-1B, open-weight and available on Hugging Face under the Apache 2.0 license. These models are not general-purpose language models; instead, they are meticulously engineered for the singular, narrow task of vulnerability localization.

In practical terms, Alongside the models, Cisco has also released the Vulnerability Localization Benchmark (VLoc Bench), a robust 500-task agentic evaluation, also under the Apache 2.0 license. This benchmark provides a standardized way to measure the effectiveness of such models in real-world scenarios.

Addressing a Critical Software Security Bottleneck

The core problem Antares addresses lies in the disconnect between abstract vulnerability knowledge (found in public databases, advisories, and Common Weakness Enumerations) and the concrete, sprawling codebases developers manage. Manually linking these two requires extensive effort:

• Searching unfamiliar code structures.
• Deciphering naming conventions.
• Tracing intricate call paths.
• Comparing numerous candidate files.

For example, Cisco emphasizes that this initial triage is where a significant portion of security costs accumulate. By automating and improving this step, Antares frees up valuable developer time and accelerates the remediation process.

Antares does not aim to replace the entire application security toolchain. Tools for dependency scanning, secret scanning, dynamic testing, container checks, threat modeling, and expert review remain essential components of a comprehensive security strategy. Instead, Antares acts as a powerful augmentation, enhancing the efficiency of existing workflows.

Understanding the Antares Architecture and Operation

The Antares family comprises three decoder-only transformer models with 350 million, 1 billion, and 3 billion parameters. All models are initialized from IBM Granite 4.0 checkpoints and share a common tokenizer and architectural features, including grouped-query attention, SwiGLU MLPs, RMSNorm, RoPE, and shared input/output embeddings.

How the Agent Loop Works

That said, Antares operates within a constrained agentic loop, utilizing a suite of three tools. The model is given only a Common Weakness Enumeration (CWE) category description, without any advisory text, file hints, or severity details.

It then interacts with a Docker sandbox (with networking disabled) by issuing read-only terminal commands. The output from these commands is truncated before being fed back into the model’s transcript.

Each task has a budget of 15 terminal calls. The model concludes by calling submit_vulnerable_files with a ranked list of file paths or submit_no_vulnerability_found if no vulnerability is detected. This agentic approach allows Antares to adaptively explore and analyze the codebase.

VLoc Bench: Evaluating Real-World Performance

Interestingly, The VLoc Bench is a comprehensive evaluation tool that draws 500 tasks from 290 unique, real-world repositories. These repositories are sourced from public GitHub Security Advisories across six major ecosystems: npm, pip, Maven, Go, Rust, and Composer. The benchmark covers 147 unique CWE categories, with a significant portion (78%) carrying assigned CVE identifiers.

The ground truth for VLoc Bench is derived directly from security patches, with modified files in the fix serving as labels (excluding tests, documentation, and configuration files). The benchmark features two phases:

1. Phase A: Evaluates the model’s ability to identify vulnerabilities in a vulnerable snapshot, scoring File F1 (a measure of precision and recall).
2. Phase B: Assesses the model’s True Negative Rate on a patched snapshot, testing its ability to avoid false alarms on fixed code.

Key Results: Task-Specific Training Trumps Scale

However, The evaluation results reveal a compelling insight: task-specific training significantly outperforms sheer parameter scale for vulnerability localization. Antares-1B achieved a File F1 score of 0.209 on VLoc Bench, surpassing much larger models like GLM-5.2 (753B parameters), which scored 0.186. Even Antares-350M (0.135 File F1) outperformed Gemma-4-31B (0.101) and Gemini 2.5 Flash (0.102).

The most powerful variant, Antares-3B (which is not yet released), reached a File F1 of 0.223, closely approaching GPT-5.5’s score of 0.229. Notably, Antares-1B also recorded the highest recall among all evaluated systems at 0.224.

Meanwhile, When compared to traditional static analysis tools run under the same evaluation protocol, Antares demonstrates a clear advantage:

• Semgrep: 0.086 File F1
• CodeQL: 0.023 File F1
• Horusec: 0.020 File F1

Cisco interprets these results as an indication that rule-based scanners, while capable of finding some vulnerable files, struggle to adaptively inspect repository context, a capability where Antares excels.

The Power of Specialized Training

In practical terms, The remarkable capabilities of Antares are not inherent in its base checkpoints. Untrained Granite 4.0 models scored negligible File F1 (around 0.000-0.001) under the same protocol, despite having tool-calling ability. The heavy lifting is performed by:

• Supervised Fine-Tuning (SFT): This phase significantly boosts performance across all scales. The SFT corpus is predominantly cybersecurity reasoning (71.5%), complemented by code search trajectories (15.4%) and general reasoning (13.1%). All reasoning traces were generated by a single teacher model, GPT-OSS-120B, to maintain consistency.
• Gradient-based Reinforcement Learning for Policy Optimization (GRPO): GRPO further enhances File F1 by 11% to 25% and, crucially, cuts run-to-run standard deviation by 42% to 65%. This reduction in variance makes Antares evaluations more reliable for applications like continuous integration (CI) scans. Rewards for GRPO are verifiable and programmatically computed, covering localization quality, submission behavior, tool-use compliance, exploration, and penalties for malformed output.

Interestingly, GRPO also influenced the learned strategies of the models. The 350M and 1B models predominantly used search commands (87-89%) and submitted more files, while the 3B model balanced search (52%) with read commands (37%) and submitted fewer files at higher precision.

Deployment and Future Outlook

For example, Antares represents a cost-effective solution, with a full 500-task sweep costing under $1 on a single H100 GPU, significantly less than other large models. This efficiency makes it a practical option for integration into development workflows.

While the Antares-3B model, which shows the strongest performance, has not yet been released, and the released models currently lack published Phase B (false-alarm) numbers, Antares-350M and Antares-1B are readily available for security researchers and developers to explore.

That said, Cisco’s Antares marks a significant advancement in leveraging specialized AI for critical software security tasks. By focusing on the precise problem of vulnerability localization and demonstrating that targeted training can outperform raw parameter scale, Antares offers a promising new tool for building more secure software.

Expert Perspective

From an industry angle, the clearest signal around AI vulnerability localization is how it may influence antares. The story reads less like a one-day spike and more like a marker of broader movement.

The next phase will depend on how quickly teams, regulators, or customers react. In practice, that gives AI vulnerability localization room to reshape expectations across security over the near term.

For readers focused on practical impact, the best next step is to watch what changes around vulnerability once attention turns into execution.

Frequently Asked Questions

Why does AI vulnerability localization matter right now?

Bridging the Gap: AI for Software Vulnerability DetectionFor readers tracking the shift, In the complex world of software development, identifying and patching vulnerabilities is a constant race.

What broader change could AI vulnerability localization signal?

Developers and security teams often face the daunting task of sifting through vast, unfamiliar codebases to pinpoint the exact location of a known flaw, a process that is both time-consuming and expensive.

What should the market watch next around AI vulnerability localization?

Cisco Foundation AI is stepping up to this challenge with the release of Antares, a new family of security-focused Small Language Models (SLMs) designed specifically for vulnerability localization.Meanwhile, Antares aims to streamline the critical first triage step in software security: given a description of a vulnerability and a repository, it efficiently identifies the files containing the flaw.

Source: https://www.marktechpost.com/2026/07/21/cisco-foundation-ai-releases-antares-350m-and-1b-open-weight-models-that-localize-known-vulnerabilities-inside-real-codebases/