CVE-2026-72670 in Fleet exposes proxy authentication credentials and private keys through the 'read agent policies' permission. A low-privilege user with this access can retrieve credentials intended only for agent runtime use, enabling lateral movement without exploiting any software flaw — the attacker simply uses valid credentials.
The CVSS 7.7 is accurate for direct impact, but the EPSS of 0.00304 is dangerously misleading. EPSS models the probability that a vulnerability will be exploited — but this CVE's operational risk lives entirely downstream of exploitation, in what an attacker does with stolen credentials. A proxy password or private key is a reusable artifact that grants access without triggering any exploit-detection signature. The probability model captures exploitability, not post-exploitation value, and credential-class exposures are systematically undervalued as a result.
More critically, credential-based access silences your detection stack. Every SIEM rule, EDR policy, and anomaly model that looks for 'unusual process plus lateral movement' becomes blind to an actor using credentials that are supposed to be there. The detection infrastructure cannot distinguish legitimate proxy use from attacker pivoting because the signal — credential authentication — is identical. This is the blast radius that no CVSS/EPSS pairing captures: not just what an attacker steals, but what defensive visibility they automatically gain.
The permission architecture itself reveals the pattern. 'Read agent policies' was likely designed when policy content meant operational rules — deployment targets, firewall exceptions. Proxy credentials migrated into that data structure over time, and the permission survived because it still worked. Nobody audited a permission that returned success. This is the 'permission archaeology' failure: security boundaries become invisible once they stop producing errors, while the data they protect quietly changes from configuration to credentials.
For defenders: audit your 'read_*' permissions for objects that have been silently promoted from low-value configuration to high-value secrets. The Fleet vulnerability is one instance of a broader class where credential-surface drift goes undetected because the permission model is a fossil — preserved in amber, still functional, never flagged for review. Prioritize credential-exposure CVEs on operational risk, not exploit probability, and assume any permission that surfaces configuration content may now surface credentials.