Sophos June 2026 Telemetry: AI Agents Trigger 56% of Credential-Access EDR Blocks via DPAPI and LOLBins

Sophos behavioral telemetry counted unique machines where Claude Code, Cursor, and OpenAI Codex reproduced attacker patterns: GStack /browse skill invoked PowerShell DPAPI decryption of browser credentials; --dangerously-skip-permissions runs enumerated Credential Manager via cmdkey; certutil and bitsadmin pivots fetched Python installers after initial blocks; and PowerShell writes to the startup folder tripped persistence rules. These actions matched high-signal detections originally tuned against human intruders.

The evidence trail reveals a collision between autonomous tooling and static EDR logic. Agents operate inside trusted user sessions with valid credentials, exactly the malware-free path CrowdStrike's 2026 report quantified at 82 percent of detections. Attacker use of the same agents, documented by Sophos one month earlier for EDR evasion testing with Claude Opus 4.5, and research on poisoned-input hijacking, means the same rule set now fires on benign automation, deliberate red-team use, and compromised agents.

Operational significance lies in the loss of behavioral signal. Enterprises scaling AI coding agents will face rising false-positive volume unless EDR vendors add context for agent processes and permission flags. Without tuning, alert fatigue will mask genuine intrusions that already favor valid credentials over dropped files.

Next quarter will test whether vendors ship agent-aware exclusions or defenders manually carve out dev-tool paths, increasing the window for living-off-the-land persistence.