Safety within the Publish-Mythos Period
Why the Fundamentals You Ignored Are the Solely Issues That Will Save You
In 2023, a colleague and I wrote a cybersecurity information for companies of any dimension. It was not glamorous work. No person was asking for an additional whitepaper about multi-factor authentication (MFA) and community segmentation. The business had heard all of it earlier than: Harden your gadgets, phase your networks, deploy endpoint detection and response (EDR), centralize your logs, check your backups, validate your designs. These usually are not revolutionary concepts. They’re the form of suggestions that get well mannered nods in consumer conferences after which get quietly dismissed someplace between finances approval and implementation.
We wrote the information anyway. Not as a result of I assumed we have been saying one thing new, however as a result of after years of incident response work, I stored strolling into the identical rooms, trying on the similar gaps, and having the identical conversations with organizations that had simply been breached. The assault vectors modified and the tooling developed, however the cause organizations obtained damage was nearly all the time the identical – the fundamentals weren’t in place. In that paper we posed questions that, when answered actually on the strategic stage, may reveal the actual state of a corporation’s defenses. We lined endpoints, networks, cloud providers, bodily safety, staffing, and logging. It was designed to be helpful whether or not you had a group of 500 safety analysts or a single IT individual sporting a number of hats.
The core thesis was that patching alone isn’t a safety technique. You want a basis that holds when patching fails – as a result of finally, patching will fail.
This situation finally arrived in April 2026.
Anthropic introduced Mission Glasswing and Claude Mythos Preview, an AI mannequin that autonomously found 1000’s of high-severity zero-day vulnerabilities throughout each main working system and net browser. Not theoretical weaknesses or potential points – working, exploitable vulnerabilities. One was undiscovered for 27 years in OpenBSD, the working system chosen particularly as a result of it’s stated to be among the many most safe on the planet. That is what occurs when vulnerability discovery stops being a human-speed exercise.
It dawned on me the whole lot we wrote about in 2023 – each suggestion, each query we posed -had simply develop into dramatically extra pressing, as pace is the brand new issue within the conventional danger triad. Cisco set out the strategic model of this argument in its Shields Up steering after working with Mythos Preview. What follows is its operational companion.
The brand new math
Earlier than Mythos and different frontier massive language fashions (LLMs), the vulnerability lifecycle had a rhythm that almost all safety groups had internalized. A researcher discovers a vulnerability, and weeks or months cross whereas an exploit will get developed. After a vendor releases a patch, organizations deploy it on their very own schedule. There was slack within the system, which gave organizations time to triage, check, and be sluggish however nonetheless survive.
After AI and LLMs, the primary two levels of that lifecycle collapsed to near-simultaneity. AI discovers the vulnerability and writes the exploit in minutes, not weeks. However the final two levels, patch launch and patch deployment, stay human-driven processes working at human pace. The hole between discovery/exploit and patch/deploy has widened from a manageable delay right into a structural hole.
The numbers make this concrete. The FIRST 2026 Vulnerability Forecast tasks a median of roughly 59,000 new CVEs this 12 months, with a 90% confidence interval reaching as much as 118,000. In 2025, 48,185 CVEs have been printed, a 21% enhance over the 12 months earlier than, which works out to roughly 131 new vulnerabilities disclosed each single day. NIST acknowledged that CVE submissions grew 263% between 2020 and 2025. Beginning April 2026, NIST introduced it might solely prioritize enrichment for CVEs showing in CISA’s Recognized Exploited Vulnerabilities (KEV) catalog, software program utilized by the federal authorities, and significant software program beneath Government Order 14028. Every little thing else goes to the again of the road.
When speaking about this knowledge in buyer briefings, I framed it round three elements: the minutes from discovery to take advantage of, the 1000’s of zero-days found, and the way AI accelerates attackers and defenders equally. The Cloud Safety Alliance was express about this of their April 2026 evaluation. The flexibility to find vulnerabilities at AI scale isn’t intrinsically a defensive functionality. It’s a dual-use functionality whose impact relies upon solely on who has entry and what constraints govern their use. We’re fortunate that frontier fashions take duty for a way they’re used, however there are a lot of open-source fashions with much less oversight.
When vulnerability administration fails, who do you fall again on?
The best way I take into consideration post-frontier mannequin protection, and the way in which I’ve been presenting it to safety leaders, follows a three-stage fallback mannequin.
The primary pillar is vulnerability administration. Scan, prioritize, patch, repeat. That is the place most organizations have concentrated their safety spending for 20 years. Patch velocity can not match AI-driven discovery charges. With 59,000+ CVEs projected for 2026 and rising, the quantity exceeds organizational capability to triage, check, and deploy (in manufacturing, dwell). Not all vulnerabilities even have patches on day zero; some are deemed as “operational danger,” or it might take years to revamp programs or {hardware}. Vulnerability administration isn’t useless, however it’s not the first line of protection; it’s now one enter amongst many. That is the place Cisco IQ turns into important. Its digital interface offers full asset visibility, safety hardening insights, and danger assessments, permitting you to proactively determine vulnerabilities and harden your programs within the face of mounting CVE volumes. Automating what you may might be key to resilience acceleration.
When patching fails, you fall again to the second pillar: the “old-fashioned” hardening that appears to be forgotten in period of EDRs. That is the place the 2023 whitepaper turns into a information:
We really helpful constructing golden pictures that incorporate applicable safety logging, refreshing them each 6 to 12 months, and making use of the newest hardening requirements. The whitepaper from 2023 asks questions that almost all organizations nonetheless can not reply confidently: Are well-known safety requirements for hardening adopted persistently throughout all gadgets? When was the final time core system golden pictures have been reviewed for weaknesses? Are golden pictures a part of safety evaluations?
The third pillar is detection and response. Hardened programs don’t stop exploitation, however make it tougher, slower, noisier, and survivable. Detection and response are what catches the exploitation that will get via, and in a post-AI exploitation world, some exploitation will get via. That is given and must be assumed.
This implies EDR, NDR, and XDR for visibility throughout layers. Behavioral detection is vital when zero-days outpace signature updates. An attacker utilizing an AI-discovered vulnerability nonetheless must execute code, set up persistence, transfer laterally, and exfiltrate knowledge. These actions produce behavioral alerts {that a} correctly configured EDR can detect no matter whether or not the precise vulnerability was beforehand recognized. It implies that we will use risk looking to search out what automation misses. It additionally means you want incident response functionality for when prevention fails. New assaults will emerge. The query isn’t whether or not you’ll be compromised. It’s now how rapidly you may detect, include, eradicate, and recuperate.
Validation isn’t optionally available
Having the precise merchandise deployed is critical, however not adequate. You additionally have to understand how they work – and right here is the place most organizations have a blind spot the dimensions of a continent.
The query each safety chief ought to be asking proper now’s “Do my controls really work? Not on paper, however beneath real-world assault situations?” Penetration testing solutions that query. So does assessing your configurations towards CIS benchmarks and hardening what falls brief. Risk modeling takes it additional by mapping the assault paths an actual adversary would use towards your particular structure, not a generic danger matrix.
Breakout assessments deserve particular consideration. They check the boundaries between community segments. Can an attacker transfer from a compromised endpoint to vital infrastructure? From IT to OT? From one enterprise unit to a different? In a post-AI world the place a zero-day can present preliminary entry to community phase, the integrity of these boundaries is arguably an important architectural property of your community. Discovering out they’re damaged earlier than an actual adversary does is the distinction between a containable incident and an existential disaster.
Then there may be the response facet, and that is the place I see the widest hole between what organizations suppose they’ve and what they really have. IR playbooks which have by no means been examined usually are not playbooks. They’re hopes. Purple group workouts are what flip these hopes into muscle reminiscence, the sort that determines whether or not your group freezes or acts when an actual incident hits. Proactive risk hunts catch what your automation missed. When the whole lot has been examined and nonetheless was not sufficient, emergency incident response is the aptitude that will get you from compromised to recovered.
The total image is a cycle. You need to stop safety points with merchandise and hardening, validate with testing and evaluation, and reply with looking and incident response – all of it backed by risk intelligence, and all of it working collectively as a system, not as disconnected level options checked off a compliance spreadsheet.
What didn’t change
AI won’t get bored with system exploitation, so danger will get realized a lot quicker than up to now. Due to this, we now add “pace” to danger equation. It turns into Threat = chance x impression x pace versus simply Threat = chance x impression. AI doesn’t change the ideas of cybersecurity. MFA nonetheless blocks credential theft; segmentation nonetheless prevents exploit cascading into the setting; EDR nonetheless detects exploitation conduct, reminiscence abuse, and makes an attempt to “write” to reminiscence segments; centralized logging nonetheless information occasions for detection and investigation; and examined backups nonetheless allow restoration.
These statements have been true earlier than any LLM/AI vulnerability discoveries, they’re true after LLM/AI, and they’re going to stay true after no matter comes after present stacks. As a result of they function at a layer of the safety stack that’s unbiased of how briskly vulnerabilities are found. They work whether or not the attacker used a recognized CVE or a recent zero-day, and whether or not the exploit was written by a human researcher over three weeks or by an AI in three minutes.
That is the structural perception constructed across the whitepaper in 2023. No person had predicted that LLM/AI vulnerability discovery explosion, however we had seen, again and again in incident response engagements, that the organizations that survived breaches weren’t those with the quickest patching cycles. They have been those that had constructed their safety foundations earlier than the breach arrived. The present AI acceleration doesn’t await finances cycles, board approvals, or strategic plans. It rewards preparation and it punishes delays.
