Antares: Foundation Models for Agentic Vulnerability Localization
Antares is a family of compact, cost-effective foundation models (350M–3B parameters) trained via a two-stage pipeline of supervised fine-tuning and reinforcement learning that achieves agentic vulnerability localization performance comparable to GPT-5.5 while significantly outperforming open-weight models over 200 times larger.
Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
Imagine you are a detective trying to solve a crime, but the crime scene isn't a single room; it's a massive, sprawling city with millions of buildings, each containing thousands of rooms. The only clue you have is a vague description of the crime, like "someone stole a specific type of lock." In the world of computer software, this city is a "codebase"—the entire collection of files that make up a program. The "lock" is a security flaw, known as a vulnerability. Finding exactly which file holds that flaw is called "vulnerability localization." It's a critical job because if you can't find the broken part, you can't fix it. Traditionally, security experts have used two main tools: rigid rulebooks (static analysis) that look for known patterns but miss new tricks, or giant, super-intelligent AI models that are incredibly smart but slow, expensive, and often require sending private code to the cloud, which companies hate doing.
Enter a new approach that treats this problem like a game of "hide and seek" played by a smart, nimble agent. Instead of just reading a book or asking a giant brain, this new method teaches a computer to actually walk through the digital city, open doors, peek inside, and follow clues just like a human detective would. The big question researchers asked was: Do you need a giant, expensive brain to do this, or can a smaller, faster, and cheaper "detective" learn to be just as good if it's trained specifically on the right skills?
The paper introduces Antares, a family of small, compact AI models designed to be these nimble digital detectives. Built on top of existing "Granite" models, Antares comes in three sizes: 350 million, 1 billion, and 3 billion parameters (think of these as the number of "neurons" or connections in the AI's brain). The researchers didn't just feed these models a bunch of security textbooks; they taught them how to act. They used a two-step training process. First, they used Supervised Fine-Tuning (SFT), which is like giving the AI a crash course in security concepts and teaching it how to use a computer terminal (the command line) to navigate folders and read files. Then, they used Reinforcement Learning (RL), a method where the AI plays the game thousands of times. Every time it successfully finds a vulnerable file, it gets a "reward." Every time it gets lost or guesses wrong, it gets a penalty. Over time, the AI learns a specific strategy: search, verify, and refine.
The results are surprisingly counter-intuitive. The paper suggests that task-specific training matters much more than raw size. In their tests, the largest Antares model, Antares-3B, achieved a score of 0.223 on a specialized test called the Vulnerability Localization Benchmark (VLoc Bench). This score is nearly identical to GPT-5.5, a massive, closed-source model that is likely many times larger and far more expensive to run. Crucially, Antares-3B outperformed substantially larger open-weight models, including GLM-5.2 (which has over 700 billion parameters). While the smaller Antares models (350M and 1B) showed impressive improvements over their base versions and outperformed some larger general-purpose baselines, they did not surpass the top-tier performers like GPT-5.5 or GLM-5.2. Instead, the key takeaway is that a compact model trained specifically for this task can compete with giants, whereas a giant model without this specific training often falls short.
The researchers also discovered that the difficulty of finding a bug isn't about how "bad" the bug is (its severity), but rather how messy the code is. They found that models struggled most in complex ecosystems like Maven (a Java tool) where files are buried deep in layers, but did much better in simpler ecosystems like pip (Python) or npm (JavaScript). Furthermore, they found that the AI's success drops sharply when a vulnerability is spread across many files or when the codebase is huge (over 10 MB). This suggests that while the AI is great at searching, it still struggles to keep track of a massive amount of evidence at once.
Perhaps the most exciting finding is about speed and cost. Because Antares is small, it can run on a single powerful computer chip (an H100 GPU) in a local office. The team calculated that running the entire 500-task test took about 15 minutes and cost less than $1 in electricity and hardware rental. In contrast, using the giant GPT-5.5 for the same job would take about 5 hours and cost roughly $141. This means companies could potentially run these security checks on their own private servers, keeping their secret code safe from the internet, without breaking the bank.
However, the paper is careful to note that this isn't a magic wand that solves everything. The AI still fails on the hardest, most complex cases where the evidence is scattered across dozens of files. The researchers suggest that the limit isn't the number of commands the AI can type, but its ability to prioritize which clues actually matter. They also emphasize that while Antares is a powerful tool for finding bugs to fix them, it is strictly intended for defensive security and should never be used to attack systems.
In short, Antares shows that you don't need a giant brain to be a good detective; you just need the right training. By teaching a small AI how to search, verify, and refine its guesses in a real-world environment, the researchers created a tool that is fast, cheap, private, and surprisingly effective at finding the needles in the digital haystack.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.