The same intrusion, with Cerberus in the loop

This is the exact frontier-lab agent intrusion — the same ~17,600 machine-speed actions — replayed through a runtime that guards the agent's tool calls and its egress. The unguarded attack ran for five days and pivoted from a sandbox into the internal network. Press play and watch where it ends instead.

2026-07-09 → 07-13 UTC ~17,600 actions replayed Contained · Day 1 · 14:11 UTC 0 records left the perimeter

2026-07-09 02:28

Day 1 / 5 · UTC

Attacker actions before interception

of 17,613 the attack would have taken

Cerberus verdict

observing

building provenance on every tool call

Blast radius

contained

agent boxed at the first outbound

Attack chain across trust boundaries

Cerberus severs the chain at egress

Phase activity

what the attack never got to do

Live action stream

tool calls + Cerberus verdicts

Volume per day

what never happened

Day 1 (07-09) - Initial access, foothold, C2 established.

Companion to the July 2026 incident disclosure. The unguarded timeline mirrors the public reconstruction; the guarded lane shows where the Cerberus gateway (fail-closed egress guard + Lethal-Trifecta detection) refuses the agent's first outbound. Cerberus guards the agent's tool-call and egress boundary inside your own environment — it does not claim to prevent the upstream model/sandbox foothold; it stops the exfiltration and every downstream pivot that needs egress. Illustrative reconstruction; indicators genericized. See Cerberus →