- Alex Stamos said attackers' AI costs are heading toward zero while defenders' costs stay high, so defenders need cheaper ways to investigate alerts.
- He argued for tiers: cheap classifiers first, then a small model, then bigger models, because frontier models are too costly for every log and every line of code.
Alex Stamos, a security expert and chief information security officer at the AI software company Cognition, argued that security is now a contest over cost. He made the case on The Verge's Decoder podcast, in the episode "AI broke the cybersecurity world. Now what?", published on 7 November 2026. His point: attackers pay almost nothing to run AI, so defenders must find much cheaper ways to respond.
What was said
Stamos said attacks are becoming agentic, meaning AI software carries them out with little human help. The AI is billed by the token, a small chunk of text. He said what attackers pay per token is heading toward zero, while defenders' costs stay high.
He said it will be very hard to make attacks expensive, so the answer is to lower the cost of defence. His example was the work of a security operations centre: watching logs, spotting an intrusion and deciding how to respond without shutting down the company.
Stamos said companies will have to replace human staff on that work with tiers of software. The first tier cannot be a large language model. It has to be "something much dumber", he said, because logs run to terabytes or petabytes. He named older tools such as traditional classifiers, which sort data quickly. They raise an alarm, a small model looks at it, and that model can call a bigger one.
He passed on a secondhand figure. A colleague at a large security platform company told him one investigation on a frontier cyber model costs $900 on their product. Stamos said that cannot work against ransomware groups whose marginal cost is zero.
Why it matters
Our reading: a security tool is only as useful as the number of alerts you can afford to examine. Stamos said even firms with access to the top models scan only their most important code. Vast amounts of older code will never get that scrutiny.
For anyone buying security software, one question follows. What does a single investigation cost, and what handles the volume before the expensive model is called? A product that needs a frontier model for every alert may not scale to the logs Stamos described.
The other side
Stamos works for a company that trains its own models, from open-weight ones, to push costs down. He said Cognition has done some of this but has "a lot more work to do." He also said very few people are focused on the problem. His case therefore comes from someone with a stake in cheaper models.
The $900 figure is secondhand and covers one product. The host pushed on the deeper worry: some cost advantage always sits with the attacker, because defence cannot be free. The excerpts do not show a full answer on how the balance settles. Stamos did say there is no way to let bots loose to do all the defence.
Written by the WebPulse Newsroom with AI assistance, and checked by our editorial review: every quotation was verified against the recording's transcript. How we use AI.
The conversation this talking point comes from
- Decoder: AI broke the cybersecurity world. Now what? (2026-11-07)
