Agent governance is getting real. This month alone, the signals are everywhere. - @Gartner_inc warned that enterprise GenAI apps will see rising security incidents as agentic AI and MCP adoption grows. - @Microsoft launched an Agent Governance Toolkit for runtime security. - @msftsecurity made the point even more clearly: identity alone cannot decide whether an action should happen now, for this user, under this business and regulatory context. - @NIST (NIST) is now treating AI agent identity and authorization as a real standards problem. - And recent research from Unit 42 showed how an overprivileged AI agent could become a path for data compromise. But agent governance is not just an identity problem. The missing layer is context security at runtime. That is because an agent is not like a human user with one fixed role. The same agent may help a marketing person in one moment. And help a sales person in the next. It may also call another agent. Or use another tool. So the real question is not just who the agent is. It is what that agent should be allowed to see, carry, share, or act on right now. So there are really two layers to solve for. First, identity. Is this really the agent it claims to be? Second, runtime context control. Given this user, this task, this data, and this workflow, what should that agent be allowed to access, preserve, unmask, share, or even act on? This is why at @protectoai, we are building CBAC: Context based access control That means detecting sensitive and confidential data in real time. Protecting it without breaking context. Allowing selective reveal only when policy allows. And carrying those controls across chained agents and tools. But the missing layer is still #AIcontext security at runtime. That is where enterprise trust in #AI will actually be won.

Why #Glasswing? LLMs can now read binary almost the way it reads code or language.  That changes cybersecurity. Anthropic says its upcoming version #Claude #Mythos Preview can reverse engineer stripped binaries into plausible source code, reason over what the software is doing, and then look for vulnerabilities from there.  So this is no longer just about scanning source code. It is about finding weaknesses directly from binaries too.  Anthropic’s public examples of such vulnerabilities found by Mythos are already notable. - 27 year old OpenBSD bug - this could have huge implication - 16 year old FFmpeg flaw that survived five million automated test hits. And Linux kernel vulnerabilities chained into full machine compromise.  That, combined with major gains in reasoning and autonomy, is why Anthropic did not broadly release the model. Instead, it launched Project #Glasswing to give defenders early access to secure critical software first.  The security ecosystem is still catching up to pace of AI innovation. The model may be ready. Is the world ready? #AI #CyberSecurity #Anthropic #ProjectGlasswing #ReverseEngineering #AppSec #Infosec