The Long Read · Ins7ghts Daily, Saturday 14 November 2026
For thirty years the scarce thing in security was the person who could find the hole. Everything downstream was built on that scarcity. Bounty programmes assume finders are rare enough to pay individually. Disclosure windows assume the next person to notice is weeks away. Quarterly patch cycles assume the queue arriving is roughly the size of the queue you cleared. All of it rests on one number: the cost of discovering a flaw.
Three pieces of work published this week say that number is falling, and they come from three different kinds of organisation.
OpenAI's security-tuned model found, exploited and patched large numbers of severe Linux kernel flaws at a dramatically lower cost per exploit than earlier models. The write-up is careful: it still struggles with hardened targets and the hardest real-world challenges. The part that matters for a buyer is the pricing. Token and caching costs now make large-scale scanning financially sensible, which turns discovery from a project into a line item. The reporting puts the shift plainly: timelines that ran in months or years compress into hours.
Anthropic's Anvil points the same capability inward. Automated red-teaming found safety-critical failures in frontier models far faster than human-only bounty programmes. The analysis is honest about what that does and does not prove: the results come from internal benchmarks and privileged access, so they are not a safety certificate, and the real measure of value is the reduction in consequential defects rather than the speed of the test. Fair. But speed is what changes behaviour, and speed is what arrived.
Then Elastic's security team took apart five agentic attack frameworks that run without asking anybody. Two details from it are worth printing. One framework blocks the obviously destructive commands, rm -rf and mkfs, and ships with an empty allow-list by default, which means shell commands can reach any host. And revoking an AWS role session does not immediately kill existing long-term access keys, with token revocation taking 60 to 90 minutes to propagate.
Hold those two facts next to each other. Finding a flaw is heading towards cents. Pulling a credential back takes an hour and a half, assuming somebody noticed and had the authority to do it.
That asymmetry is the structural read. Security spending for a decade went into detection, because detection was the bottleneck. Detection is becoming abundant on both sides of the fence at the same time, which means the bottleneck moves to the part nobody was buying: the change window, the approval queue, the owner who can authorise a revocation at 2am, the service that cannot be restarted in business hours because finance reports run on it. None of those is an AI problem. All of them are now the limit on how fast you can respond.
The rest of this piece is for subscribers.
Ins7ghts Daily is free. One letter a morning from Yves Mulkers on what moved in data and AI, and why it matters to your side of the table.
Already subscribed? Sign in at the top of the page.