Securing Inherited AI Agents Within Enterprise Platforms
Platform-native agents bypass traditional security by inheriting existing application configurations and permissions.
1The Agentforce Vulnerability Chain
Standard features like Email-to-Case create unintended pathways for unauthenticated outsiders to interact with internal AI agents. Data flows from public email to case records, triggering agent prompts and outbound replies. This architecture is not a misconfiguration but a documented feature, exposing internal agent instructions to external users without any security intervention.
2Platform-Level Security Gaps
Security risks emerge from the host application rather than the AI model itself. Procurement-stage vendor reviews fail to identify these threats because agents are often deployed post-procurement. Prompt filters remain unaware of the data source, as they lack visibility into the underlying platform intake mechanisms like Web-to-Case or email integrations.
3Shadow Agents and Permissions
Shadow agents proliferate as employees build tools to automate tasks like pipeline reporting or invoice reconciliation. These agents inherit the creator's full system permissions, effectively multiplying individual access rights across the enterprise. Unlike manual shadow AI, these agents run unattended, creating a massive, invisible attack surface that standard security scanners cannot detect.
4The Failure of Static Governance
Enterprise governance often relies on acceptable-use policies and monthly committees that lack operational enforcement. These documents fail to track how many agents are active, what data they access, or who owns them. The OWASP Agentic Skills Top 10 highlights this 'No Governance' category, noting that third-party platform agents remain invisible to external scanners.
5Automated Agent Proliferation
Modern orchestration platforms allow coding agents to deploy sub-agents without human oversight. Each sub-agent gains its own connections and permissions, bypassing traditional review processes. Governance models assuming human-initiated requests are becoming obsolete as autonomous systems scaffold and deploy new agents, creating a complex, unmonitored web of internal automation and data access.
6Operational Evidence for Compliance
The EU AI Act imposes severe penalties, reaching up to 35 million euros or 7% of global turnover. Compliance requires operational evidence, not just paper policies. Historical precedents like British Airways and Marriott demonstrate that regulators penalize the inability to prove operational control, making bottom-up governance essential for meeting modern regulatory standards.
7Implementing Bottom-Up Governance
Effective governance requires an operational test at every layer of the software stack. Organizations must maintain a live inventory of agents, their owners, and their specific configurations. Platforms like Reco provide visibility by discovering agents across multiple vantage points, including those bypassing SSO, allowing security teams to assign risk profiles before applying policies.
8Securing the Agent Ecosystem
The City-Forum campaign, active since March 2025, proves that attackers exploit the gap between platform features and actual enterprise configurations. Security teams must treat agent configurations as governable assets. By shifting from static policy documents to operational visibility, enterprises can control permissions and reduce insider risks before these configurations lead to breaches.
Original article: The Hacker News — Securing Inherited AI Agents Within Enterprise Platforms