Thinking about TARA in the AI era.
How do we even “assess” attacker expertise now, when anyone may be a prompt away from being a “professional hacker”? What does the PS5 hacking story tell us? And why is a car not just about ECUs and items?
asintsov.com/notes/2026-11-1…
AI pushed new CVEs to 12x last year's pace. Breaches didn't move. Why? I modeled it as a forest fire: vulns are just lightning, most fires stay small, a few giants do the damage.
AI decides how often lightning strikes. The forest decides how much burns.
asintsov.com/notes/2026-10-0…
Let me say that again...
You store pointers at the _destination_ address of a memcpy.
You glitch during memcpy ().
You get that pointer into PC.
No, it's not sci-fi. It's the "instruction corruption" fault model. And we pioneered that.
See thread below 1/N.
This attack showed that the data at the destination of a copy can be abused just like the data at the source.
We had to improve this attack quite a bit as it simply took too long to get a successful glitch. The details for this optimization will be explained during our training.
Application Security and Vulnerability Assessment getting a significant advantage from GenAI (context-driven knowledgebase). That helps security teams understand the root cause of the problem faster and significantly reduces the latency in producing security fixes at scale.
"... detected several remotely exploitable bugs in AMI MegaRAC BMC"
"... whole attack sequence: from having zero knowledge about a remote AMI BMC with enabled IPMI (yeah, right) to flashing a persistent firmware implant to the server SPI flash"
Looking forward to this talk!
Also found interesting, that ChatGPT works much better if you ask to use LangSec approach: translate logic into grammar, and input as a language and try to find a Weird Machine, works more efficient at my example than just "check the pseudocode/logic for security issues"