Post #4149596

2026-07-27 23:04 UTC

@mttaggart@infosec.exchange I am left with an open question though: when the subsidies end, does this remain a dangerous technology that is actually cost-effective? e.g. would a nation-state attacker with very deep pockets be better off just hiring a lot of people, or running an LLM?

Replies (2)

@mttaggart@infosec.exchange 2026-07-27 23:12

@glyph@mastodon.social Oh I think it remains dangerous. We know open weight models can find vulns just fine. https://aisle.com/blog/aisle-matches-anthropic-mythos-on-freebsd-zero-days And let's assume that capacity maps roughly onto offensive capability. So between rapid discovery of new vulnerabilities, exploitation of existing ones, and orchestration of attacks, the danger doesn't go away. Also when the subsidies end, a collapse will mean a collapse in the price of hardware for local compute. I really think we'll have to deal with the consequences of unleashing this poison for some time.

Open ##4149595
@sharkfie@infosec.exchange 2026-07-27 23:59

@glyph@mastodon.social @mttaggart@infosec.exchange I had okay results using it to build "agents" for blue team stuff and others on my team, but it's the situation where I often wonder if it's the most efficient (and deterministic) way or if a LLM is necessary. I don't think costs are subsided much with AWS etc. but I could be wrong. I'm not a AI booster, but the incentives and budget are for everyone to use AI so I try to make do.

Open ##4150445