AI Is Spitting Out Working Exploits for $3.61. Microsoft Says the Old Playbook Is Dead.
A senior Microsoft security chief told Black Hat USA that AI tools have made finding and weaponising software flaws so cheap and fast that the entire industry's approach to defence needs to change, starting now.
Key points
- Microsoft's internal AI scanning tool found roughly 200 Linux kernel vulnerabilities and automatically generated working exploit code for 182 of them.
- The average cost of finding a vulnerability and building an exploit was $3.61, with a turnaround time of 21 minutes per flaw.
- Microsoft's Security Response Center is now processing and patching twice as many vulnerabilities every six weeks compared to earlier this year, nine times the volume it handled in March 2025.
- A senior Microsoft engineer warned that automatically generated exploits will be "pretty commonplace" across the industry by the end of 2025.
- Switching to memory-safe programming languages, such as Rust, is one of the most effective steps vendors can take to cut the number of exploitable flaws at their root.
Software vulnerabilities have always existed. The buffer that defenders once relied on, the gap between a flaw being discovered and a criminal building a weapon from it, is shrinking fast. Microsoft is making that argument loudly.
What exactly did Microsoft announce?
At Black Hat USA last week, David Weston, group manager on Microsoft's Windows security team, delivered a keynote titled The End of Rare: Defending When Offense Is Cheap. The message was not subtle.
Weston described Microsoft's MDASH tool, short for Multi-model Agentic Scanning Harness, a system of AI assistants that reads software code, hunts for security flaws, and tries to build a working attack against each one automatically, without a human in the loop. We covered MDASH's earlier capability jump on 28 July, when Microsoft paired a new cybersecurity-only AI model with GPT-5.4 and scored 95.95% on the CyberGym benchmark at half the previous cost.
Pointing MDASH at Microsoft's internal Azure Linux distribution produced roughly 200 previously unknown vulnerabilities in the Linux kernel, the core software layer that manages a computer's hardware. Of those, MDASH automatically built crash-triggering proof-of-concept exploits, test attacks that prove a flaw is real and usable, for 182 of them. Many went further, becoming fully working attacks capable of handing an attacker root access, meaning complete administrative control over the machine.
The average cost per vulnerability found and exploited: $3.61. Average time: 21 minutes.
"These are serious vulnerabilities, the kind that I used to take a year to bespoke craft," Weston said. "They're being spit out at industrial speed."
| Metric | Figure |
|---|---|
| Linux kernel flaws found by MDASH | ~200 |
| Exploit-level proofs generated automatically | 182 |
| Average cost per flaw detected and exploited | $3.61 |
| Average time to generate an exploit | 21 minutes |
| Vulnerability volume increase since March 2025 | 9x |
| Google's Android memory-safety flaw rate, 2019 | 76% |
| Google's Android memory-safety flaw rate, 2025 | Below 20% |
Should ordinary people be worried?
Not in a panic-stations way, but the trend matters. The more easily criminals can find and weaponise software flaws, the faster companies need to patch their products, and the shorter your window of safety between a flaw existing and someone actively exploiting it.
Weston said Microsoft's Security Response Center is now processing and patching twice as many vulnerabilities every six weeks as earlier in the year, nine times the volume from March 2025. That workload increase tracks directly with rising AI tool use across the security research community. It is not a Windows-specific problem: Linux, Android and every other major operating system is seeing the same acceleration.
What does Microsoft say defenders should do differently?
Stop playing whack-a-mole, essentially. Weston's argument is that reacting to one vulnerability at a time cannot keep pace when AI lets attackers find new flaws faster than defenders can fix old ones. "We don't wanna go vulnerability for patch," he said. "Hand-to-hand combat with attackers will cause us to lose in defense."
His preferred alternative is switching to memory-safe programming languages, coding tools designed so that an entire category of common mistake simply cannot happen. Rust and Golang are the main examples. Around 70% of the vulnerabilities that major vendors patch today stem from memory-safety errors, the kind that safer languages eliminate by design.
Google's experience with Android is the clearest proof point. After prioritising Rust for new Android code, Google reduced memory-safety vulnerabilities from 76% of all Android flaws in 2019 to below 20% in 2025. Microsoft itself rewrote its Azure hypervisor, the software layer that keeps separate virtual machines isolated on a shared physical server, in Rust and deployed it across 1.5 million virtual machines without incident.
A US government research agency, DARPA, is running a project called Tractor that automates converting older C code, a widely used but memory-unsafe language, into Rust. Microsoft Research's own tool, RustAssistant, uses AI to spot and suggest fixes for Rust coding errors.
What about the limits of just rewriting code in a safer language?
Valid question. Yan Shoshitaishvili, an associate professor at Arizona State University, gave a companion keynote at the same conference that poured cold water on the rewrite-everything-in-Rust idea.
His team used AI-assisted tools to study OpenHarmony, the open-source software base underpinning parts of Huawei's HarmonyOS for mobile devices. They found dozens of flaws covering Bluetooth attacks, device takeovers and location-data leaks, discovered faster than his team could responsibly write them up and report them.
When Shoshitaishvili tested a Rust rewrite of core Linux utilities, memory-safety bugs largely disappeared. Logic-level vulnerabilities, design flaws in how the software behaves rather than how it handles memory, reappeared in new forms. Safer languages are a powerful tool, not a complete cure.
Common questions
Does this mean criminals already have these AI exploit tools?
Not necessarily the exact same tools Microsoft built internally, but the underlying AI models are widely available, and independent researchers have shown similar results. Weston's point is that the barrier to building such a pipeline is now low enough that well-resourced criminal groups almost certainly have something comparable.
What should employees or ordinary users actually do right now?
Keep software and devices updated promptly. If a company you use announces a security patch, the window between "patch released" and "criminals actively exploiting the flaw" is getting shorter, so treating updates as optional is riskier than it used to be.
