Cisco's Antares AI model ranks risky files for security teams, scores 0.823 on source but 0.000 on binaries
shashib · x · 2026-10-16
- Cisco Foundation AI's Antares is a family of small models that act like a sniffer dog for bug hunting: feed it a vulnerability description (CWE-labeled) plus a copy of a codebase, and it searches like a human investigator — spotting patterns, reading suspicious files, dropping dead-end leads — returning a ranked file list with its reasoning trail.
- Cisco's own tests show a stark gap: the strongest version scores 0.823 on readable source code but 0.000 on compiled binaries, i.e. it completely fails from an attacker's view.
- 190 of 500 real bugs in the test set beat every model tested, showing hard limits of small models on real-world vulnerability triage.
- The analysis draws on the author's conversation with Cisco chief AI scientist Amin Karbasi at Black Hat USA and the group's September results posted October 6.
More from Models
- Meta open-weights Muse Spark Open: 816B total/42B active, 1M context, coding-focused — PyTorch · 2026-10-16
- NVIDIA's Nemotron 3.5 Super VL: 120B MoE multimodal model with 5x agentic throughput on one GPU — markjeffrey · 2026-10-16
- Polymarket gives Meta just 3% to lead AI by year-end; Anthropic at 44% — Polymarket · 2026-10-16
- Muse Spark Open agent details: subagents, computer use, 1M-token project context — togethercompute · 2026-10-16
- Together AI serves Muse Spark Open serverless for agentic coding and multimodal workloads — togethercompute · 2026-10-16
- Muse Spark Open goes live on Together AI as a Day 0 launch partner — togethercompute · 2026-10-16