How AI-accelerated exploitation compresses the defender’s window from weeks to minutes, why patching faster is no longer a strategy on its own, and where FileDNA Content Analysis, Disarm and Reconstruction changes the math on file-based delivery.
For roughly three decades, defenders won a specific race more often than they lost it. A vendor shipped a security fix, the clock started, and defenders scrambled to deploy the update before an attacker could reverse-engineer it into a working exploit. The reason they usually won was simple and durable. Turning a patch back into a reliable exploit was slow, specialized, expert-level work that generally required weeks, and sometimes months, of scarce reverse-engineering talent. That labor cost was the moat. It bought defenders time to stage rollouts, honor uptime commitments, and run regression tests without leaving every unpatched system exposed for long.
That moat is gone. Recent research from Anthropic’s Frontier Red Team, amplified in a widely read analysis at The Hacker News, shows that frontier language models have collapsed the single most time-intensive step in an N-day campaign from expert-weeks into a matter of minutes. The industry shorthand for a known, disclosed, but not-yet-patched vulnerability has always been the N-day. The uncomfortable new reality is that the meaningful unit of time is closer to the N-hour.
Where FileDNA fits from the outset: The exploit is only half of any real campaign. The other half is delivery, and file-based delivery remains the workhorse of intrusion. FileDNA Content Analysis, Disarm and Reconstruction (CADR) does not depend on knowing which vulnerability an attacker weaponized or how fast they did it. It rebuilds every inbound file to a known-good, threat-free state at the ingress boundary, stripping the weaponized payload out of the document, archive, or script before it ever reaches a user or an unpatched endpoint. When the patch gap shrinks to zero, the delivery layer becomes the layer defenders can still control.
The patch itself is the roadmap
There is a hard asymmetry at the center of N-day exploitation that a lot of security programs never internalize. The patch meant to protect you is the same artifact that arms the attacker against you. The moment a vendor ships a fix, the difference between the old code and the new code, the diff, tells anyone watching exactly what was broken and precisely where. Skilled adversaries compare the pre-patch and post-patch binaries or source, a practice called patch diffing, locate the changed logic, and reverse-engineer the vulnerability the patch was written to close. A working exploit against every system that has not yet updated then becomes, historically, a matter of time.
The incidents most defenders remember reflected that older, slower timeline. WannaCry struck fifty-nine days after Microsoft shipped MS17-010 in 2017. The public exploit for Citrix Bleed in 2023 took roughly two weeks to appear. In Mandiant’s 2020 study of N-days, sixteen of twenty-five vulnerabilities took a month or more to weaponize. Those weeks were the defender’s oxygen. They were also, it turns out, an artifact of how expensive human reverse engineering used to be.
Reverse-engineering a patch used to take weeks. Now it can take an hour.
Anthropic’s red team measured the new timeline directly, and the numbers are stark. Working against Mozilla’s Firefox, the team handed a frontier model nothing but a public patch diff and two builds, one vulnerable and one fixed. Across eighteen recent Firefox JavaScript-engine patches, Claude Mythos Preview autonomously produced eight working, code-execution exploits. Its first exploit landed in under an hour after Mozilla shipped the patch. The Firefox release actually carrying that fix to users was still eighteen days away. The exploit existed before the update did.
Firefox was close to a best-case scenario for defenders, which is what makes the result so sobering. Firefox updates itself automatically, adopts fixes with a simple browser restart, ships urgent one-off releases when needed, and has tightened its point-release cadence from monthly to roughly weekly. If the patch gap is exploitable even there, it is exploitable almost everywhere else, where enterprise remediation routinely stretches into weeks or months.
The closed-source results were harder still, and arguably more alarming. Against twenty-one Windows kernel vulnerabilities, with no source code available and only stripped binaries plus decompiler output to work from, the model built proof-of-concept crashes for eighteen of them, the fastest arriving in thirty-one minutes. It then chained eight of those all the way to full SYSTEM privilege, at a cost of roughly two thousand dollars each in API credits. One of those SYSTEM-level chains targeted a bug Microsoft had rated “Exploitation Unlikely.” That severity rating, like most in the industry, is calibrated to human researchers. That calibration no longer holds.
Crucially, this is not a single locked capability behind one gated model. The publicly available models, even with their safeguards engaged, produced working exploits too. Fewer of them, but enough to prove the capability has diffused. As Anthropic’s own team put it, N-hour is closer to the reality we now operate in.
Why speed of exploitation raises the value of the delivery layer: As the cost of weaponizing any given patch falls toward zero, the scarce resource for defenders is no longer reverse-engineering talent on the attacker’s side. It is any control that works independently of patch state on the defender’s side. A weaponized exploit still has to arrive somewhere, and for a large share of intrusions it arrives inside a file. FileDNA CADR treats every inbound file as untrusted by default and reconstructs it to a clean baseline, so a malicious macro, an embedded exploit, or an obfuscated loader is removed during reconstruction regardless of which fresh N-day the payload was built to trigger downstream.
You cannot patch your way out of this
The instinctive response to a collapsing patch gap is to patch faster. The data says that instinct loses. Verizon’s 2026 Data Breach Investigations Report put the median time to fix a known-exploited flaw at forty-three days, up from thirty-two the prior year, with only about a quarter of such flaws ever fully patched. Even top performers close only thirty to forty percent of known-exploited vulnerabilities in the first week. Meanwhile, tracking from the Zero Day Clock places 2026’s average time-to-exploit at under twenty-four hours, down from roughly fifty-three days in 2024.
Put those two curves on the same chart and the problem is obvious. Exploitation has accelerated into hours while remediation has, if anything, gotten slower. Patches wait for regression testing, change windows, and uptime guarantees. Taking production offline to outrun an exploit is just a different kind of outage. With roughly one hundred thirty-five new vulnerabilities disclosed every day, up about forty percent year over year, no team is clearing its backlog. When everything in the queue scores a near-maximum severity, the queue effectively prioritizes nothing. Today’s breaches increasingly happen inside precisely that gap.
The more useful question is not “what is vulnerable?” It is “what can an attacker actually exploit here, would our controls stop the attempt, and can we prove it?” That reframing moves the conversation from patch counts to demonstrated control effectiveness, and it is where a defense-in-depth layer that does not depend on patch state earns its place.
The delivery step still has to happen, and files are how it happens
Anthropic’s researchers are careful to note that turning an exploit into a full intrusion still requires additional work. Delivery, targeting, and evasion all take effort that the model did not automate. That caveat is easy to read as reassurance. It is better read as a map of where defenders should concentrate, because delivery is the step defenders can most directly disrupt.
For a large and stubborn share of real-world intrusions, delivery means a file. A spreadsheet with a malicious macro. A PDF carrying an embedded exploit. A ZIP or ISO smuggling a loader past a mail gateway. A booby-trapped document that triggers a memory-corruption bug in the very parser or browser engine an attacker just weaponized from a fresh patch diff. The exploit may now be produced in an hour, but it still has to ride into the environment on something, and files remain the most reliable vehicle.
How FileDNA CADR breaks the delivery step: Detection-based defenses ask a fragile question, “have I seen this before?” That question fails against a brand-new N-hour exploit with no signature and no reputation history. CADR asks a fundamentally different question, “what is the known-good version of this file?” FileDNA analyzes each inbound file, disarms active and executable content, and reconstructs a clean, structurally valid version that preserves the legitimate content the user needs while discarding everything that does not belong. There is no waiting for a signature, no reliance on prior knowledge of the exploit, and no dependence on the endpoint having been patched. A weaponized document is neutralized during reconstruction at the ingress boundary, before the freshly minted exploit inside it ever gets the chance to fire.
Validation, and where a patch-independent layer belongs in it
The security industry’s emerging answer to the N-hour problem is exposure validation. Rather than assuming exploitability from a severity score, defenders prove it. Where firing a live exploit is safe, autonomous penetration testing runs the real chain against reachable assets and delivers the strongest possible evidence. That approach only reaches an estimated ten to fifteen percent of a typical environment, though, because live detonation is off-limits on business-critical systems, restricted networks, and air-gapped segments, which tend to be exactly the assets that matter most, and because a brand-new disclosure has no safe public exploit on day one.
For the remaining eighty-five to ninety percent, defenders decompose a vulnerability into its chain of attacker techniques and validate each link against their actual controls, endpoint policy, segmentation, allow-listing, and firewall rules. Break a required link, and the exposure is not exploitable in that environment, even on assets that can never be touched and against threats no one has weaponized yet. Continuous testing of the live prevention and detection stack against current techniques closes the loop, revealing what gets blocked, what slips through silently, and where a control has drifted. The discipline becomes a repeating cycle of validate, decide, fix, and re-validate, which turns a critical finding into a defensible decision to patch, mitigate, monitor, or accept, rather than a guess anchored to a severity number.
CADR as a validated, break-the-chain control: Exposure validation asks whether a required link in the attack chain can be broken with evidence. For the very large class of intrusions that begin with a file, FileDNA CADR is that break. When the delivery link is a weaponized document and reconstruction removes the payload before it reaches the target, the chain fails at ingress no matter how quickly the exploit was produced or how far behind the endpoint’s patch level has fallen. That is a control that holds on exactly the assets live pen-testing cannot safely touch, air-gapped, restricted, and business-critical systems, and it holds on day one of a disclosure, before any safe exploit exists to test against. It converts patch speed from a life-or-death race into one input among several, because the file-borne path in has already been closed.
What CADR does not claim
Precision matters here, and overclaiming helps no one. FileDNA CADR is a file-boundary control. It neutralizes threats carried inside files as they cross the ingress boundary into an environment. It is not a substitute for patching, and it is not a memory-safe rewrite of your kernel. It does not disarm an exploit that a defender has already staged locally outside its inspection path, and it does not protect against attack vectors that never touch a file at all, such as a purely network-level exploit fired directly at an exposed service. CADR is a defensive layer, and its power comes from being combined with patching, exposure validation, segmentation, and detection, not from replacing any of them.
What it does claim is meaningful and, in an N-hour world, increasingly decisive. Where a threat arrives as a file, FileDNA removes the weaponized content before it can act, without needing to recognize the specific exploit, without waiting for a signature, and without depending on the destination system being patched. As the patch gap collapses toward zero, that patch-independence is the whole point.
The board question has changed
Vendors are already moving to shrink the patch gap. Mozilla tightened Firefox’s release cadence. Microsoft ships urgent fixes out-of-band and, increasingly, as reboot-free hotpatches. The most durable long-term fix attacks the supply of bugs directly, by migrating critical components to memory-safe languages like Rust and by hardening code with mitigations that retire whole exploit classes at once. All of that is worth doing, and none of it will finish this year.
In the meantime, the question a board asks its security leaders has shifted from “are we patched?” to “are we secure right now, and can you prove it?” Patching remains necessary, but on its own it is no longer sufficient, because the timeline that made patch-and-pray viable has evaporated. The defenders who come through the N-hour era in good shape will be the ones who layered controls that do not depend on winning a race they can no longer win. For the very large fraction of attacks that still arrive as a file, that layer is FileDNA CADR, neutralizing the payload at the boundary, on every asset, before the exploit inside it gets its hour.
Sources: Anthropic Frontier Red Team, “Measuring LLMs’ impact on N-day exploits” (June 2026);
The Hacker News, “N-day is Becoming N-Hour. Patching Faster Won’t Save You.” (July 2026);
Verizon 2026 DBIR; Zero Day Clock.