What happened
On Oct 22, 2026 Composio (agent tool-access platform spanning 1,500+ apps) introduced 'Firewall,' a hook layer that evaluates each agent tool call before execution and each returned tool result after execution, using tuned small language models (e.g., Jev) plus a built-in prompt_injection_check handler. It lets developers write natural-language conditions (e.g., block sharing internal pricing outside acme.com) and is documented as experimental on paid plans.
Why it matters
Indirect prompt injection arrives inside tool results (emails, PRs, web pages) rather than the chat box, so screening the tool boundary is a distinct, increasingly necessary control. Composio placing this at the agent-to-tool integration layer gives it broad reach across the agent ecosystem, though block rates and effectiveness are not yet independently measured.
Applicability
Developer and security platform teams building agentic workflows on Composio should evaluate the Firewall hooks for temporary, least-privilege guardrails; teams with strict data-retention needs should weigh that zero-data-retention mode drops session context and weakens chained-behavior checks.