6 posts in news
The next agent breach won't be a prompt — it'll be a permission
Non-human identities now outnumber humans by as much as 144 to 1, and most sit outside the governance perimeter entirely. As AI agents get their own credentials and act on behalf of many users, the industry is converging on intent-based authorization: not who is the identity, but is this agent, acting for this user, allowed to do this specific thing right now?
Intertrace Threat Research — Threat intelligence
When agents talk to agents: the A2A blast radius
Google's Agent-to-Agent protocol, backed by 50+ partners, lets independent agents discover and delegate to each other — and delegates credential and card verification entirely to implementers. Researchers have already catalogued impersonation, unauthorized registration, recursive denial-of-service, and malicious-artifact attacks. Multi-agent systems multiply autonomy; they multiply the attack surface with it.
Intertrace Threat Research — Threat intelligence
EchoLeak: the first zero-click AI exploit, and what it changes
CVE-2025-32711 turned a crafted email into silent data theft from Microsoft 365 Copilot — no click, no attachment, no user action. It is the clearest proof yet that an LLM assistant with read access to your data is an exfiltration path. Here is the mechanism, and where a runtime gateway breaks the chain.
Intertrace Threat Research — Threat intelligence
The MCP supply chain became an attack surface faster than anyone secured it
A critical RCE in MCP infrastructure, a config-swap exploit against Cursor, and a backdoored gateway package downloaded 47,000 times in three hours. The Model Context Protocol connected agents to everything — including the attacker. A field guide to the incidents, and how to mediate MCP instead of trusting it.
Intertrace Threat Research — Threat intelligence
Indirect prompt injection has gone operational
OWASP still ranks prompt injection as the #1 LLM risk, and its own researchers now say it is unsolved. Meanwhile findings landed against Slack AI, Microsoft 365 Copilot, Cursor, and GitHub MCP, and a 2026 survey reported 88% of organizations hit a confirmed or suspected AI-agent incident. The proof-of-concept era is over.
Intertrace Threat Research — Threat intelligence
August 2, 2026: agent governance stopped being optional
The EU AI Act's enforcement powers for general-purpose AI took effect on August 2, 2026, and NIST opened a dedicated AI agent standards initiative in February. Both point at the same requirement: controls that reach runtime, logs that survive review, and an inventory that includes your agents. Here is what changes and how to be ready.
Intertrace Threat Research — Threat intelligence
Need something else?
Email support@intertrace.ai with topic requests, or try the Simulation Lab and the documentation hub.