Two hours of keyword searching on a 300-file project is what Julio Gomez says one new security advisory costs a software team, and the bill repeats for every repository. Antares, a small, free artificial intelligence model from Cisco, reads the advisory and ranks the files in your codebase to open first.

Gomez, a Principal Engineer in Cisco's Office of the Chief Technology Officer, spent nine hours on an Apple laptop running Antares-1B against 500 cases from Cisco's own benchmark, and he wrote it up on October 22 as Part 1 of a five-part series. His accuracy score of 0.168 against Cisco's published 0.209 works out to about 80 percent by my arithmetic (Gomez; Cisco Foundation AI).

The check my July note asked for came from inside Cisco

In July I wrote in a technology note for Info-Tech Research Group that no outside lab had repeated Cisco's results, and I told security teams to test the model on their own code. Gomez has now done a version of that and posted his results on GitHub so anyone can rerun them. He works at Cisco, so the test shows the numbers can be repeated. A lab with no stake in the outcome has yet to run it.

Gomez says so himself. He calls Cisco's comparison table a vendor benchmark and limits his claim to one run on one laptop. Cisco's own score averages three runs on much faster data center machines.

One right file on the list counts as a win

For daily use, Gomez points to how often the list contains at least one file with the flaw. In his test that happened in 31.7 percent of the cases, about one in three (Gomez). The list matched the fixed files with nothing extra in 8.6 percent (Gomez). Flip the first number and you get the miss rate: about two cases in three came back with no flawed file on the list.

A ranked list with a record of the search attached can carry more authority in a meeting than it has earned.

Cisco tested each case twice, once with the flaw present and once after the fix. Gomez reports that the model tends to name files even when the code is already fixed. On your screen, a list that missed the flaw looks the same as one that found it.

The savings are largest where accuracy runs out

Cisco's technical report puts numbers on the limit. On projects under 100 kilobytes, Antares-1B scores 0.843, and in the 10-megabyte tier and above, GPT-5.5 overtakes it (Cisco Foundation AI). Cases with five or more vulnerable files see performance collapse, and the model does better at finding one foothold than at recovering every file involved. The search runs on a budget of 15 commands per case, which Gomez says cannot cover a 100-megabyte codebase.

The biggest, oldest systems take the longest to search by hand, and they are where Antares loses ground. I expect a codebase that has grown for a decade or more to fall outside the safe zone.

The cost advantage holds for the small cases. Cisco's report lists $0.71 for the Antares-1B run across all 500 cases against $141 for OpenAI's GPT-5.5 (Cisco Foundation AI).

The biggest, oldest systems take the longest to search by hand, and they are where Antares loses ground.

Five posts lead the reader into Cisco's free toolkit

Gomez gives Antares the job of showing where to look and Foundry Security Spec the job of proving the flaw is real. Foundry is Cisco's free blueprint for checking AI-flagged problems before anyone acts on them. CodeGuard, Cisco's set of security rules for AI coding assistants, handles prevention. SecurityBrief reported in May that Foundry lays out a process from first detection through final reporting, and that CodeGuard supplies the rules for the detection step.

• October 2025: Cisco open-sources CodeGuard.
• May 2026: Cisco releases Foundry Security Spec.
• July 21, 2026: Cisco releases Antares-350M and Antares-1B.
• October 22, 2026: Gomez publishes Part 1 of his series, with Part 2 on Foundry due the following week.

In my July note I argued that the free release matters more than the test score because it pulls developers toward Cisco's way of working. Part 1 fits that reading. It ends by sending you to Foundry, and Part 2 promises a run on popular open-source projects with every finding checked against the actual code. You get a documented process with no license fee and pay in staff time to put it to work.

Antares starts from a small free model that IBM released in October 2025 for anyone to use and modify. A vendor review that treats Antares as one company's work misses IBM's part in it.