Wiki › Engineering with agents

AI Vulnerability Tsunami

The expectation that AI makes exploitation faster than defenders can patch, pushing defence toward isolation, agentic remediation, and a stricter baseline.

The "vulnerability tsunami" (also called the "vulnpocalypse") is the expectation that AI models will make finding and exploiting software flaws so much cheaper and faster that defenders can no longer patch in time. Defence then has to shift towards isolation, automation and a stricter baseline.

The warning

On 20 August 2026 Jason Clinton, deputy CISO at Anthropic, wrote that "the vulnerability tsunami is coming in just a few months and there won't be enough time to patch critical infrastructure". He called for network isolation, including air gaps and data diodes, "now". His post quoted the NSA's announcement of a joint advisory on AI-assisted attacks against industrial controllers.¹

That advisory, AA26-231A (NSA, CISA, FBI, DOE and EPA, 18 August 2026), reports active reconnaissance and capability development against internet-exposed Siemens S7 PLCs in US manufacturing, energy, water, chemical, food and commercial facilities. The attackers locate the devices through scanning services such as Censys and ZoomEye. Their exploitation scripts are AI-generated from public information and disguised as legitimate monitoring tools. Recommended mitigations:²

• inventory all PLCs and apply patches;
• block S7comm (TCP 102) at the perimeter;
• put a DMZ between IT and OT networks;
• use unidirectional gateways for data historian links "where appropriate";
• allowlist engineering workstations and enable PLC passwords;
• monitor and hunt for anomalies.

Clinton's emphasis goes beyond the advisory. For the agencies, isolation is one item among many. For him it is the main point: industrial equipment patches slowly, so it should not be reachable at all.

Signals since

• Practitioner report (October 2026). Alan Wells describes a week "on the front lines of the vulnpocalypse" fighting an active attack. He credits defensive-access frontier models (Anthropic's Mythos, OpenAI's Daybreak) and an agent pipeline for the shift to proactive patching. See Agentic Security Remediation.³
• Breach statistics. A practical security guide citing the 2026 Verizon DBIR says that vulnerability exploitation (31% of breaches) has overtaken stolen credentials (13%) as the top initial access vector for the first time. The same guide says median remediation time rose from 32 to 43 days, and only 26% of known exploited vulnerabilities were fully fixed. These figures come via a secondary source and have not been checked against the report here.⁴

Defensive responses

1. Reduce reachability. Air gaps, data diodes, segmentation, and nothing industrial facing the internet. This is the critical-infrastructure answer when patching cannot keep up.¹

2. Match the attacker's speed with agents. Defenders use frontier models and agent loops to find, rank, patch and verify faster. See Agentic Security Remediation and Agentic Software Factories.³

3. Raise the baseline. Remove the cheap paths, such as reused passwords, unpatched devices and unverified payment requests, so commodity automation has less to hit. See Baseline Security Controls.⁴

Caveats

The "tsunami" framing comes largely from people at frontier labs that also sell defensive cyber access, so the warning and the remedy share a commercial interest. The timeline ("a few months") was a forecast. The advisory documents AI-assisted reconnaissance and tooling, not a confirmed wave of successful infrastructure attacks. This page is marked contested until broader evidence accumulates.