To truly understand the unprecedented Cybersecurity crisis looming in front of us, we need to understand something simple: For the most part AI hasn’t made attackers smarter (yet), it has made them easier and cheaper. And that’s important because it changes the risk math for every company.
At Oracle AI World in Las Vegas this week, Sam Altman was asked what companies should do with AI in the short term. His answer was not about productivity gains or deploying AI agents. He said that if a company does one thing, it should be to understand the security threat landscape from AI and work to mitigate it. A week earlier at Stanford he mentioned that he “no longer thinks we’re going to be able to just patch all the bugs” (The Next Web). And this was after he, along with another 100+ technology leaders, signed a letter warning that the window to defend against AI-driven attacks is narrowing.
Many have argued this is an attempt at free publicity or gaining some leverage over inevitable AI regulation efforts, but when the CEO of a company selling a technology tells customers to worry about the technology it sells, it is worth digging deeper. This is not a post about AI taking over the world. It is about something simpler: how the cost of finding and attacking a vulnerable company has collapsed, and why that turns old, “tolerated weaknesses” into potentially catastrophic issues. In other words, the “typical” risk equation for vulnerability management has fundamentally changed for everyone.
How we used to decide what was “low risk”
Every security program I have worked with rates risk in roughly the same way: likelihood times impact. Abstracting away the many different ways of implementing a risk assessment, impact is generally easier to agree on. Likelihood on the other hand is where many judgment (i.e. subjective) calls inevitably happen in spite of serious efforts by many in the community to remove subjectivity from the process.
For years, one underlying assumption was made (explicitly or implicitly) when risk-ranking several weaknesses: “exploiting this issue takes effort”. This assumption is why so many organizations carry a long tail of known, unpatched issues:
- the legacy internal application nobody wants to touch because it still runs payroll
- the VPN appliance two firmware versions behind
- the web server with an old library that would need a full regression test to upgrade
- the forgotten test environment still reachable from the internet
Each of these was marked “low” or “accepted” for an arguably defensible reason at the time: finding the flaw required skill, and turning it into a working exploit required even more skill (and time). This line of reasoning would then conclude that a skilled attacker has better and possibly easier targets than a 200-person manufacturer or a regional accounting firm. In other words, “obscurity and effort” were mitigating controls, even if nobody formally acknowledged them as such.
What has changed in 2026
Simply put, AI has made our underlying assumption obsolete. Finding flaws is now not only cheap, but incredibly fast.
In April, Anthropic announced Claude Mythos Preview, a model the company claims it “chose” not to release publicly because of how effectively it finds vulnerabilities. Among its findings was a flaw in OpenBSD that had gone undetected for 27 years. Let’s think about this for a second. OpenBSD is an operating system known for its focus on security. Then by late May, scans of more than 1,000 open-source projects had produced 23,019 candidate findings, an estimated 6,202 of them high or critical, while only 97 upstream patches existed. Of course, OpenAI and others have built comparable tools as well.
The consequences are easy to see in the public data. The US National Vulnerability Database (NVD) recorded 45,207 flaws between January and late July 2026, almost as many as in the entire year in 2025 (a record-setting year). Microsoft alone disclosed 642 bugs in July, nearly five times the same month a year earlier, and one Chrome update fixed 433 issues against 11 in the equivalent release the year before (Bloomberg via The Star).
More concerning, the average time to exploitation has shrunk from weeks to hours. Recorded Future puts the average time for attackers to exploit a vulnerability at 24 hours in 2026, down from 72 hours last year (see Bloomberg report above). Synack’s 2026 State of Vulnerabilities report describes the gap between discovery and exploitation as hours. ZeroFox’s summary of the 2026 Verizon Data Breach Investigations Report puts it at roughly 10 hours, with exploitation of vulnerabilities now the most common way attackers get in (31% of breaches). This last part is significant because in the past it used to be easier to exploit “human weaknesses” via social engineering.
With the attack mechanisms being increasingly automated as well, an adversary can now gain access to your systems and data much faster. In May, Sysdig documented an intrusion driven by an AI agent that went from code execution on an exposed notebook server to a dump of an internal database in under an hour, with no prepared playbook (ZeroFox).
And the consequences are real. In August 2025, Anthropic reported that a single criminal had used its coding agent to automate reconnaissance, harvest credentials and break into at least 17 organizations, including healthcare providers, emergency services and government bodies. The AI helped decide which data to steal and wrote the ransom notes. Demands ran into six figures: One attacker, seventeen victims. No “advanced persistent threat” or “nation-state threat actor” required.
A sober view of the numbers
Experienced practitioners know very well that finding vulnerabilities was never the hardest part (although it is much easier/faster now). The truly hard part is fixing them. When Anthropic announced all the vulnerabilities Mythos had found, they also mentioned that more than 99% of those vulnerabilities were still unpatched (Fortune).
So the change is not that every bug is suddenly being exploited. The challenge now lies in who gets targeted, and how cheaply.
Before AI-powered attacks, an attacker needed days of skilled work per target, so they chose targets carefully: large companies with high-value data and known weak spots. Now, an AI agent can scan, assess and attempt exploitation across thousands of organizations at a very low cost, so targeting becomes opportunistic. Your organization is no longer attacked because it’s “interesting”, it’s attacked simply because you are reachable and something you run is vulnerable.
That is why being “small” or “uninteresting” no longer protects anyone. A mid-sized company that reasoned “we are not a target” was arguably at least partly right for a long time. It is not right anymore.
Why this is now an existential issue, not an IT issue
For many organizations, a successful attack no longer just means “a bad week”. It is now a question of survival. Consider that:
- Extortion does not need encryption. Modern groups often skip ransomware entirely and simply steal data, then threaten to publish it or notify customers and regulators. Good backups do not help against that.
- Smaller margins means smaller buffers. A large enterprise can absorb weeks of disruption, legal costs and lost contracts. A 50- to 500-person company often cannot, especially if a key customer contract or an insurance policy depends on demonstrating reasonable security.
- The “accepted risk” is now on the record. If the breach came through a known issue that was documented and deferred, the conversation with customers, insurers and regulators becomes much harder.
What to do about it: pragmatic steps
The good news is that none of this requires a large or even a new budget to begin. But it does require a shift in priorities and it requires revisiting decisions made under assumptions that no longer hold:
- Know what is exposed. Build an accurate inventory of everything reachable from the internet: domains, subdomains, VPNs, remote access, cloud services, forgotten test systems. Attackers will build this list about you whether or not you have one.
- Re-score your accepted risks. Pull up the risk register and every exception that was approved because exploitation was “unlikely” or “difficult.” Those likelihood ratings were based on attacker effort. Re-rate likelihood under the assumption that a capable attacker can find and exploit the issue in a day (because one can).
- Patch internet-facing systems in days, not months. Prioritize anything on CISA’s Known Exploited Vulnerabilities list and anything exposed to the internet. Monthly patch cycles were designed for a world where exploitation took weeks.
- Shrink the attack surface. Retire what you do not need. Put administrative interfaces and legacy applications behind a VPN or zero-trust access rather than on the open internet. Every system you remove is one less thing to patch.
- Assume someone gets in, and limit what they can take. Phishing-resistant MFA, least-privilege access, network segmentation and monitoring for large data transfers matter more than ever when the goal is data theft and extortion.
- Test recovery and rehearse extortion. Restore from backup for real, not on paper. Run a tabletop exercise in which someone threatens to publish your customer data in 72 hours: who decides, who calls legal, who talks to customers?
- Use the same tools defensively. The AI that finds flaws in other people’s code can find them in yours. Use it to review your own applications, configurations and cloud accounts before someone else does.
- Ask the same questions of your suppliers. Your managed service provider, your SaaS vendors and your software suppliers carry their own tail of “low-risk” issues, and their access to your environment is part of your attack surface.
The question for leadership
The shift here is not only (or even primarily) technical. For years, the gap between “we know about this vulnerability” and “we have fixed it” was tolerable because exploiting it took time and skill. That gap is now where the risk sits.
The question every leader or risk executive should be asking is not “are we secure?” It is: “Which known weaknesses have we accepted because they were hard to exploit, and are we still comfortable with each of them now that they are not?”
If the honest answer is “we don’t know,” that is the place to start.
Sources
- The National: OpenAI’s Sam Altman urges companies to focus on AI security threat (27 Oct 2026)
- The Next Web: Sam Altman says AI means we can no longer just patch all the bugs (26 Oct 2026)
- Dayton Daily News (NYT): OpenAI and 100 others warn that window to defend against AI attacks is narrowing (June 2026)
- Fortune: Anthropic’s Mythos and the real problem of fixing, not finding (13 Apr 2026)
- eWeek: Anthropic Mythos finds 23,019 vulnerability candidates as patching lags (May 2026)
- Bloomberg via The Star: AI finding twice as many cyber flaws in 2026 as in 2025 (28 Jul 2026)
- Help Net Security: Synack 2026 State of Vulnerabilities report (18 May 2026)
- ZeroFox: From CVE to breach in under an hour (17 Jun 2026)
- Engadget: Anthropic admits its AI is being used to conduct cybercrime (27 Aug 2025)
