Cisco Scans 1.8 Billion Lines as AI Accelerates Vulnerability Discovery

Key Takeaways

• A major networking provider used frontier models to scan 1.8 billion lines of code across 25 languages during its initial rollout, reporting a false-positive rate below 3%.
• Project Glasswing indicates that AI can dramatically increase vulnerability discovery, shifting pressure toward prioritization, remediation, and containment.
• Modern security architectures promote layered defenses, machine-speed response, secure AI development, and governance across the AI agent lifecycle.

Perfect prevention was already an unrealistic security objective. Frontier AI is making the limitations of that strategy considerably harder to ignore.

Cisco has tested a range of frontier models against 1.8 billion lines of code across its portfolio, covering 25 programming languages during the initial scanning phase. According to the regional cybersecurity field CTO, equivalent work using earlier methods would have taken roughly eight years. The engineering team used specialized harnesses to improve fidelity and reported a false-positive rate below 3%.

That changes the economics of vulnerability research. It does not, however, automatically solve the harder operational problem: deciding what to fix, assigning ownership, testing changes, and deploying them without disrupting production.

Anthropic's initial May 2026 update on Project Glasswing brought that problem into focus. In roughly one month, participating organizations identified more than 10,000 high- or critical-severity vulnerabilities. Some teams saw their bug-finding rates jump by more than 10x. Mozilla fixed 271 issues in a single Firefox release, while Cloudflare pulled 2,000 findings with a false-positive rate better than its human testers.

The scale is not limited to technology vendors. The Government of Alberta used Claude to review 466 million lines of code across 27 ministries in about 20 hours. That example shows how quickly AI-assisted analysis can reach large and varied software estates, including public-sector environments where legacy applications and complicated dependencies can slow remediation.

Increased discovery rates do not necessarily translate to reduced risk. If remediation capacity stays flat while discovery rises sharply, security teams accumulate a larger and more accurately documented backlog.

The Verizon Data Breach Investigations Report provides relevant context. Its 2026 findings placed vulnerability exploitation first among initial access vectors at 31%, ahead of credential abuse for the first time. Only 26% of CISA Known Exploited Vulnerabilities were fully remediated, while the median time to fix reached 43 days. AI-assisted discovery could widen that gap unless organizations improve prioritization and response.

The recommended security architecture centers on an assume-breach posture rather than dependence on a single preventative control. That includes mapping defenses against MITRE ATT&CK, limiting lateral movement, placing runtime protection closer to workloads, and automating detection, triage, and containment. Phishing-resistant MFA, least privilege for people and AI agents, and accurate asset inventories remain central. So does retiring end-of-life technology before automated systems make its weaknesses easier to locate and exploit.

Cisco is also assembling open resources for different stages of the problem. Foundry Security Spec provides a model-agnostic blueprint for building auditable scanning harnesses. Project CodeGuard embeds secure-by-default rules into AI coding-agent workflows. Foundation-Sec-8B adapts Llama-3.1-8B using cybersecurity material, while Antares uses small, open-weight models to localize known vulnerabilities without sending source code to the cloud.

DefenseClaw addresses AI agent governance by scanning skills and MCP servers, inspecting prompts and tool calls, pausing risky actions for human review, and exporting evidence into existing security systems. The proprietary LLM Security Leaderboard evaluates model behavior under single-turn and multi-turn attacks. Together, these projects reflect the broader governance principles described by the NIST AI Risk Management Framework, particularly the need to measure, monitor, and manage AI-related risks throughout deployment.

The IQ platform is positioned as the commercial layer connecting continuous visibility, exposure prioritization, adaptive assessments, and infrastructure services. Still, tooling is only part of the equation. Can an organization contain an AI-assisted exploit chain when a patch is unavailable? The practical response involves segmentation, compensating controls, temporary feature restrictions, and closer monitoring until remediation is safe.

The strategic shift is straightforward, even if execution is messy. Vulnerability discovery is moving toward machine speed. Security programs now have to bring prioritization, containment, and recovery closer to that pace, while accepting that remediation capacity will remain finite.