Nice summary. Something I find interesting as an AI person that the cyber experts I know focus so much on the ecosystem and trends that existed before, where AI is an accelerant. On my side of the world, AI people love to debate if it’s caused by open models or if trusted access programs are good enough and things that actually seem much more surface level.
You’re spot on, for many of the seasoned cyber practitioner community we’ve already been seeing many of these trends and situations historically but AI is essentially accelerating and compounding them dramatically while also straining the institutions we rely on (such as NIST’s NVD etc.) as well as internal security team and tooling capacity to absorb the acceleration
I feel like closed models crossing thresholds before open is literally the best warning we could get. Open models will come, even if some companies (kimi) intentionally slow cyber capabilities. There are too many people building models, and too much economic pressure to fast follow OpenAI/Anthropic as the fastest growing businesses ever. SOMEONE will take the risk, as they have plausible deniability if someone else uses the model for harm, not them.
(The companies hosting abliterated models on APIs are ngmi imo)
The Cairn campaign here is the same NorForge report I wrote up a few days back, the $50-a-hack one. What stuck with me wasn't the exploitation numbers, it was the funding numbers sitting right next to them. The same quarter this hit, $6.8 billion went into cybersecurity, and a good chunk of the AI-security rounds specifically funded tools that map agents a company already runs internally. None of that catches a decade-old deserialization bug in a public-facing web app nobody's watching. The debt-collector framing you land on is right, the collector's going after the boring external surface nobody budgets for, while the money keeps chasing the newer, flashier category instead. The fundamentals argument at the end is correct too, the open question is why fundamentals keep losing that budget fight.
Averaged over 2,250 tenants, 2,900 sign-in attempts a minute is barely more than one per tenant per minute, which sits under most lockout thresholds. Catching that means correlating across tenants or by source, so I'd check whether any of your alerts look past a single account.
Nice summary. Something I find interesting as an AI person that the cyber experts I know focus so much on the ecosystem and trends that existed before, where AI is an accelerant. On my side of the world, AI people love to debate if it’s caused by open models or if trusted access programs are good enough and things that actually seem much more surface level.
Hoping to write about all this this weekend too!
Appreciate the read and comment.
You’re spot on, for many of the seasoned cyber practitioner community we’ve already been seeing many of these trends and situations historically but AI is essentially accelerating and compounding them dramatically while also straining the institutions we rely on (such as NIST’s NVD etc.) as well as internal security team and tooling capacity to absorb the acceleration
I feel like closed models crossing thresholds before open is literally the best warning we could get. Open models will come, even if some companies (kimi) intentionally slow cyber capabilities. There are too many people building models, and too much economic pressure to fast follow OpenAI/Anthropic as the fastest growing businesses ever. SOMEONE will take the risk, as they have plausible deniability if someone else uses the model for harm, not them.
(The companies hosting abliterated models on APIs are ngmi imo)
The Cairn campaign here is the same NorForge report I wrote up a few days back, the $50-a-hack one. What stuck with me wasn't the exploitation numbers, it was the funding numbers sitting right next to them. The same quarter this hit, $6.8 billion went into cybersecurity, and a good chunk of the AI-security rounds specifically funded tools that map agents a company already runs internally. None of that catches a decade-old deserialization bug in a public-facing web app nobody's watching. The debt-collector framing you land on is right, the collector's going after the boring external surface nobody budgets for, while the money keeps chasing the newer, flashier category instead. The fundamentals argument at the end is correct too, the open question is why fundamentals keep losing that budget fight.
Averaged over 2,250 tenants, 2,900 sign-in attempts a minute is barely more than one per tenant per minute, which sits under most lockout thresholds. Catching that means correlating across tenants or by source, so I'd check whether any of your alerts look past a single account.