Defending at Machine Speed: How Security Teams Have to Change

Security operations centres are built around a triage funnel: alerts arrive, humans assess them, a subset becomes incidents. The funnel assumes attacker activity unfolds slowly enough for that assessment to matter. Once an intrusion can move from initial access to exfiltration inside a working day — the tempo implied by the Anthropic disclosure — the funnel becomes a bottleneck rather than a filter.

Automate containment, not judgement

The instinct is to automate detection. That is usually wrong, because bad detections automated at scale generate noise faster than analysts can absorb it. The higher-return move is to automate the reversible containment actions that follow a high-confidence detection: disabling a session token, isolating a host, forcing reauthentication, revoking an API key. Each is cheap to undo if wrong and expensive to delay if right.

Automate the actions you would be comfortable reversing. Keep humans on the ones you would not.

Where the leverage actually is

• Identity hygiene: short-lived credentials, phishing-resistant MFA, and elimination of long-lived service account keys. Most automated intrusion depends on credentials that should not have existed.
• Egress visibility: exfiltration is the one step an attacker cannot skip, and it is the least well instrumented in most environments.
• Asset truth: automated attackers enumerate faster than defenders inventory. An accurate, current asset list is a security control.
• Detection as code: version-controlled, tested detections with measured false-positive rates, rather than a growing pile of untested rules.

Using AI on the defensive side

The same capability that compresses attacker labour compresses defender labour. Alert enrichment, log summarisation, first-pass triage narratives and detection authoring are all tasks where a model produces a useful draft in seconds. The discipline required is to treat model output as evidence rather than verdict: every automated conclusion should carry the raw telemetry that produced it, so an analyst can disagree quickly.

The measurement that matters

Most teams report mean time to detect and mean time to respond. Under machine-speed intrusion the more useful metric is the ratio of attacker dwell time to defender decision time. If an attacker completes a phase in twenty minutes and your escalation path takes ninety, no amount of tooling improves the outcome. Measure the gap, then close it by removing approval steps rather than by buying products.

None of this requires accepting the most dramatic readings of AI risk. It only requires accepting that the cheapest part of an attack just got cheaper, and that defensive processes calibrated to the old cost are now mispriced.