In February 2025, CrowdStrike titled its annual threat report with a warning that would have sounded speculative just two years earlier: “AI Accelerates Adversaries and Reshapes the Attack Surface.” By spring 2026, that warning looks more like a status update. Automated attack tools powered by artificial intelligence can now discover and exploit software vulnerabilities faster than most organizations can deploy patches, and the gap between discovery and defense is widening. The contest playing out across enterprise networks and government systems is no longer just between rival hacking groups or nation-states. It is between machine-speed offense and human-speed defense, and the machines are pulling ahead.
Federal agencies and vendors are treating this as operational, not theoretical
The Cybersecurity and Infrastructure Security Agency launched a pilot program for AI-enabled vulnerability detection that uses artificial intelligence to scan codebases at scale. The logic behind the pilot is blunt: if defenders can use AI to surface flaws quickly, attackers can do the same, and whichever side acts first controls the outcome. CISA has not yet published performance benchmarks from the pilot, and the agency has not publicly specified which AI techniques the program employs. But its decision to move the technology into a production-grade test environment signals confidence that AI-driven code review works well enough to matter.
CrowdStrike’s report provides the private-sector counterpart. Drawing on intrusion data from thousands of enterprise clients, the company’s 2025 Global Threat Report framed AI as a force multiplier that shortens the timeline from vulnerability discovery to weaponized exploit delivery. Adam Meyers, CrowdStrike’s senior vice president of counter-adversary operations, told reporters at the report’s launch that adversaries were already using generative AI to accelerate social engineering and to rapidly iterate on exploit code. The company tracked a record number of new adversaries in 2024 and noted that average breakout time for interactive intrusions had dropped to 48 minutes, with the fastest observed at just 51 seconds.
On the policy side, the White House issued a presidential directive in January 2025 titled “Removing Barriers to American Leadership in Artificial Intelligence.” The order revoked earlier AI governance requirements and stated its intent to remove barriers to AI development and deployment across the federal government. It did not prescribe specific defensive tools, but it signaled that the administration views rapid AI integration, including in cybersecurity, as a national priority.
The National Institute of Standards and Technology published research in January 2024 cataloguing distinct attack categories that manipulate AI system behavior: evasion attacks, data poisoning, and privacy-based exploits, each targeting a different stage of a model’s lifecycle. That taxonomy established the vocabulary federal agencies and vendors now use when discussing AI-specific threats. It also highlighted a dimension that gets less attention: AI systems themselves can become attack surfaces, not just tools wielded by attackers.
Taken together, these moves show that both government and industry now treat AI as a central factor in cyber offense and defense. CISA is experimenting with machine-scale code review. The White House has signaled its intent to accelerate AI adoption. NIST has mapped how AI itself can be compromised. And frontline incident responders say adversaries are already exploiting these capabilities in the field.
Where the evidence thins out
The strongest claims about AI-assisted attacks in 2026 rest on directional evidence rather than published benchmarks. CISA’s pilot has not released quantitative metrics comparing AI-driven scanning to traditional human-led code review. Without those numbers, the assertion that AI finds flaws “faster than humans can patch” draws on logical inference and vendor reporting rather than controlled measurement.
CrowdStrike’s threat report describes accelerating adversary timelines, but the underlying incident data remains proprietary. The company has not published granular case studies showing specific AI-discovered vulnerabilities exploited before patches became available. That is typical for vendor intelligence: it reflects real-world observations from a major incident-response firm, but it is not independently audited research. Readers should weigh it alongside federal assessments rather than treating either source as the complete picture.
The relationship between the White House AI order and actual defensive improvements is similarly unclear. No public reporting has documented which agencies adopted new AI-driven defensive tools as a direct result of the January 2025 directive, or whether the order changed patch timelines for federal systems. Policy intent and operational change are different things, and the distance between them has not been measured in any available source.
NIST’s attack taxonomy, while authoritative, is now more than two years old. The agency has not released an updated assessment reflecting how those attack categories have evolved. Newer techniques, such as large-scale abuse of generative models for automated phishing or for chaining vulnerabilities across complex cloud environments, may fall outside the categories NIST initially defined. The absence of a refreshed federal classification means the 2024 taxonomy remains the best available reference point, even as the threat landscape it describes has continued to shift.
There is also an open question about how broadly AI-accelerated exploitation has spread beyond the most capable threat actors. Nation-state groups and well-funded criminal organizations likely have the resources to integrate AI into their tooling. But the proliferation of open-source AI models, including fine-tuned variants with fewer safety guardrails, is lowering the barrier for mid-tier and opportunistic attackers too. Without public case studies that attribute specific intrusions to AI-driven discovery rather than traditional scanning, it is difficult to quantify how much of the observed increase in attack speed is genuinely new versus a continuation of long-running automation trends.
What this means for organizations patching in real time
For security teams trying to gauge their own exposure, the most useful approach is to separate what is clearly happening from what is plausible but unproven. It is established that AI can accelerate code analysis, that government and industry are racing to operationalize those capabilities, and that sophisticated adversaries are experimenting with the same tools. It is plausible, but not yet empirically demonstrated in public data, that AI has already compressed the average window from vulnerability disclosure to widespread exploitation.
That distinction matters for planning. Organizations that wait for definitive proof before adjusting their defenses risk falling behind a threat curve that moves at machine speed. But organizations that chase every vendor warning without scrutinizing the evidence risk misallocating resources on speculative threats while neglecting fundamentals.
The practical middle ground starts with the basics that have always mattered: accurate asset inventories, rapid patching processes, network segmentation, and continuous monitoring. On top of that foundation, security leaders are beginning to explore AI-assisted defenses, tools that can triage vulnerability scan results, prioritize patches by exploitability, and flag anomalous behavior faster than human analysts working alone. Google’s Mandiant division, Palo Alto Networks, and Microsoft have all expanded AI-driven detection offerings in the past year, joining CrowdStrike in a market that is growing as fast as the threat it aims to counter.
The contest between machine-speed offense and machine-speed defense is real, and it is accelerating. Its outcome is not predetermined. But the organizations most likely to stay ahead are the ones treating this shift as an operational reality right now, not waiting for a cleaner data set to confirm what the early evidence already suggests.
More from Morning Overview
*This article was researched with the help of AI, with human editors creating the final content.
