As AI models become more sophisticated, their ability to autonomously find and exploit vulnerabilities is increasing exponentially, making them a powerful weapon in the hands of cyber attackers. And with enterprises accelerating their adoption of GenAI and autonomous agents, vendors are shifting from traditional detection models toward AI-native security architectures.
This week, Microsoft, Cisco and OpenAI have each unveiled new initiatives aimed at addressing growing concerns across the enterprise market about how to secure these increasingly autonomous AI systems before attackers exploit them.
The announcements point to a transition away from static security tooling and toward agentic defense systems that can continuously evaluate and address threats across software environments.
As Cisco pointed out in announcing its initiative:
“The operating model of cybersecurity has fundamentally shifted. As frontier AI models create a new dual-front challenge, attackers are now identifying vulnerabilities at machine speed, leaving security teams struggling to keep pace with manual, legacy processes.”
Microsoft Pushes Multi-Agent Cyber Defense
Microsoft has introduced a new agentic security system, multi-model agentic scanning harness (MDASH), which combines more than 100 specialized AI agents to detect vulnerabilities across Windows infrastructure. It uses frontier LLMs including Anthropic’s Claude Mythos, OpenAI’s GPT-5.5-Cyber and others.
According to the company, the platform helped Microsoft researchers “find 16 new vulnerabilities across the Windows networking and authentication stack—including four Critical remote code execution flaws in components such as the Windows kernel TCP/IP stack and the IKEv2 service.”
Microsoft fixed the flaws in its weekly Patch Tuesday update pushed out to Windows devices.
The vendor claims that the system outperformed Anthropic’s Claude Mythos Preview and OpenAI’s ChatGPT 5.5 in benchmarking of real-world vulnerabilities.
Taesoo Kim, Vice President, Agentic Security at Microsoft, wrote in the blog post announcing the system:
“AI vulnerability discovery has crossed from research curiosity into production-grade defense at enterprise scale, and the durable advantage lies in the agentic system around the model rather than any single model itself.”
Discovering security flaws using AI is becoming an engineering problem, Kim added. It requires composition that no single prompt can achieve and must include validation to fix the flaws uncovered.
Microsoft’s system absorbs improvements in AI models, so that the targeting, debating, deduplication, and proof stages do not need to be rewritten each time there is an update. Instead, the vendor changes a configuration and re-runs an A/B test, and the customer’s investment, including per-project context, scan plugins, and proving agents, carries over.
“This is the architectural property that matters most over time, because the model lottery is going to keep playing out, and any system whose value is gated on a particular model is a system that has to be rebuilt every six months,” Kim noted.
The architecture is intended to emulate collaborative human security teams, with agents specializing in reasoning, exploit validation, triage and remediation workflows.
Microsoft is framing the initiative as a response to the widening speed gap between attackers and defenders. In parallel, the vendor also detailed work using AI-generated synthetic attack logs to improve detection engineering and training datasets for security operations teams.
MDASH is helping Microsoft’s engineering teams improve security outcomes using generally available AI models and is being tested by customers as part of a limited private preview.
Cisco Addresses the Limits of AI Security Analysis
Cisco has taken a different approach, releasing an open-source framework called the Foundry Security Spec. Rather than introducing a standalone product, the company published a model-agnostic and stack-agnostic reference architecture for building auditable AI-driven security evaluation systems.
Cisco warned that simply using AI to attempt to find and fix flaws is not enough. Omar Santos, Distinguished Engineer, AI Security Engineering, S&TO, stated:
“Organizations are investing in AI-assisted security and getting back hallucinated findings, false positives at scale, and no coverage signal.”
When security teams point an LLM at a repository and ask it to “find the bugs,” they are often given “a wall of unbounded, unverifiable output that mixes sharp insights with hallucinated findings, with no way to know what was missed or when you’re actually done,” according to Santos.
