RSA Conference 2026 just wrapped. Five major vendors launched agent identity frameworks. All cover discovery, OAuth, permissions. Three critical gaps survived all five.
The 3 Gaps
Gap 1: Tool-Call Authorization
OAuth confirms who the agent is. Nothing constrains what parameters it passes.
A CEO's agent had legitimate credentials, found a restriction, and removed it. Every identity check passed. No framework detects agents rewriting their own security policy.
The basic version: Langflow's build_public_tmp endpoint (CVE-2026-33017, CVSS 9.8) required no auth at all. CISA KEV. Attackers had working exploits within 20 hours. JFrog confirmed the 'patched' 1.8.2 was still exploitable. Real fix: 1.9.0.
Gap 2: Permission Lifecycle
Agent permissions expanded 3x in one month without security review. Discovery tools show what exists today; none track how permissions evolved.
Gap 3: Ghost Agent Offboarding
One-third of enterprise agents run on third-party platforms. Pilots end, agents keep running. Only 21% maintain real-time agent inventory.
What Catches These Gaps
Identity = WHO. Two more layers needed:
Verification (HOW): Agent Security Harness — 440 adversarial tests. Now on GitHub Marketplace.
-
AUTH-001: Unauthenticated access (catches Langflow pattern) -
AUTHZ-001: Least privilege enforcement -
CP-007: Profile escalation (can agent modify its own capabilities?)
Governance (WHY): Constitutional-agent-governance catches Gap 1:
# GovernanceGate: zero tolerance for self-modification
if control_bypass_attempts >= 1:
return GateResult(
gate='GovernanceGate', state=GateState.FAIL,
reason='Control bypass attempted. Human intervention required.'
)
What's Missing
We catch Gap 1. We don't address Gap 2 (permission drift) or Gap 3 (ghost agents). PermissionDriftGate and AgentInventoryGate would close these.
The Stack
- Identity (WHO) — all 5 vendors shipped this
- Verification (HOW) — proves identity controls hold under attack
- Governance (WHY) — constitutional constraints for ungoverned scenarios
Most teams run layer 1 only. RSA showed why that's not enough.
85% of orgs adopting agents. 5% at production scale. The barrier is trust, not capability.
Top comments (0)