The gap between a vulnerability’s disclosure and its exploitation used to be measured in days. Last month I watched it collapse to hours inside my own infrastructure. When PHP2Shell went public, attackers were already scanning, and reports of active exploitation arrived before the official patch did. One of our legacy systems runs PHP. We activated our security incident response plan the moment we learned of the risk, before upstream had even published a fix. We isolated the exposed PHP components immediately and started manual patching ahead of the official release. Remediation was complete within hours, with no material impact on the service. We then spent days on forensics, credential rotation, and log audits. The investigation found no evidence of a breach in our system.
Once the immediate fire was out, I looked up from our logs at what else had landed while we were heads-down. Our week was not an isolated event, but part of a broader wave across October: critical vulnerabilities in PHP, Node.js, Ruby, and GitLab, plus a poisoned PyPI package. This week’s Alpine OpenSSH authentication bypass is the latest entry, and the most telling part is how OpenSSH responded. In the 10.6p2 release announcement, the maintainers said they now consider long embargoes dangerous and mainly helpful to attackers, because they have recently seen AI agents rediscover the same bugs on their own, sometimes within hours. When the team maintaining some of the most security-critical software on the planet abandons the old disclosure playbook, the tempo change is no longer a prediction. It is policy.
That October wave points to a fundamental shift that most people have not realized yet: AI has collapsed the skill floor for exploit development. As IEEE Spectrum reported, the researcher who weaponized PHP2Shell used AI to turn a bug report dismissed as “just a crash” into remote code execution in two weeks. Attacks that used to require the top 0.1% of experts and months of work now need far less. Every “low priority” bug sitting in every backlog is loaded ammunition now, and the grace period between disclosure and exploitation is gone.
The problem is that most enterprise security programs are still built around that grace period. Organizations rely on daily CVE scans, weekly patch windows, and quarterly upgrade cycles. That entire model was designed for known threats moving at human speed, assuming teams have days to triage and schedule maintenance. When attackers exploit flaws before the patch even exists, a weekly patch cadence is just a schedule for discovering you were breached.
Going through the PHP2Shell response forced me to rethink what realistic defense requires now. Defense has to move from handling known threats to operating as if unknown ones are already inside. Some of what we have put in place, for reference: stricter least-privilege access across the board; agent-assisted diagnosis and analysis, because without it the speed gap between attack and defense can never be closed; agent-assisted DevOps that can patch and deploy faster than any human on-call rotation; and monitoring not just CVE feeds but the source repositories of our critical components, analyzing potential risk before any advisory exists. One honest note from our own practice: our systems carry baseline hardening, including security-hardened self-compiled components, and that genuinely helps against unknown threats. But hardening cannot stop very generic vulnerabilities, and on public-facing services the risk is especially severe.
We are still exploring more effective strategies for unknown risks. But this battlefield is already unbalanced. If your organization has not reacted to this shift yet, there is a real chance your servers are already compromised and you simply do not know it.
The old question was “how fast can we patch?” The new question is “what still works when we are too slow?” Security used to be about building higher walls. Now it is about assuming the wall is already breached and designing for what comes after. Prepare before it rains. It is already raining.