THE Hacker News piece highlights a growing blind spot in AI security: third-party agents that come embedded inside enterprise software, often without a formal adoption decision or clear governance. In a landscape studied for the 2026 State of Agent Security Report, about 1,280 products embed AI, with roughly 282 behind single sign-on and many remaining invisible to identity tooling by default.
The article argues that traditional first-party security controls assume an intentional choice and a gatekeeper; agents, by contrast, arrive inside existing applications and can act with broad reach without ever being “adopted” in a recognisable way.
A notable example cited is Salesforce’s Slack Code, launched in August 2026, where a coding agent can read context, write code, and open a pull request, yet security controls may not instrument this autonomous actor as it operates through chat-channel membership and production access.
The piece emphasises a framework of four practical questions to assess any agent: Identity, Permissions, Connectivity, and Activity. It argues that risk often lies in the scaffolding and ecosystem around the model rather than the model itself, with reach and privilege extending through connected systems. The article notes that large buyers—like JPMorgan Chase—now scrutinise agents as supply-chain risks, advocating for identity by default and deliberate entitlements.
Regulators, including the EU AI Act, are pushing inventory and oversight. A move towards live, continuously refreshed visibility—such as Reco’s graph that maps identities, permissions, and agent actions in real time—is proposed as the viable path for governance as agents proliferate, regardless of origin. It also points to recent exploits and CVEs, e.g., Citrix NetScaler CVE-2026-88772, as reminders of the evolving threat landscape.