AI Security in Development: The Agent Became an Actor
Срез по безопасности AI в разработке: где агент становится риском и где он уже ловит уязвимости
Episode participants
A solo episode without invited guests.
What we discussed
A scheduled Research Insights Made Simple episode walks through the September 2026 snapshot on AI security in software development: 153 sources checked against primary texts and three risk surfaces — the code the model writes, the tools it calls, and the agent itself with identity, permissions and network access.
The conversation follows the 2025–2026 timeline: prompt-injection records across five coding agents, four worm waves in the package registry, secrets leaking through context, two containment incidents and the identity layer the market answered with.
The central question is what is already proven on the defense side, where vulnerability discovery hit the wall of fixing, and which six bets a team can make before the next snapshot in spring 2027.