Day 2 — Identity Attacks + Voice Social Engineering

Course: SEC5xx — Detecting and Responding to AI-Generated Adversary Content
Day: 2 of 5 · ~6 hours instruction + 2.5 hour lab + breaks
Prerequisite: Day 1 (Detector's AI Stack + AI-generated phishing)
Version: 2.1 — rebuilt from 2026 threat data · see the evidence annex

Why this day was rebuilt

The v1 curriculum spent this day on deepfake audio/video detection. The 2026 evidence says the daily problem lives elsewhere: 82% of detections are malware-free (CrowdStrike GTR 2026), credential abuse appears in 39% of breaches (DBIR 2026), and voice phishing is now the #2 initial infection vector at 11% of intrusions — the top entry into cloud environments (M-Trends 2026). The adversary's AI shows up here as fluent pretexts and cloned voices aimed at help desks, not as cinema-grade video. So Day 2 now trains the identity-first kill chain end to end, and synthetic media takes its correctly-sized place: one module, anchored to the help-desk workflow where it actually bites.

By end of Day 2, students leave with:

  1. A working detection set (Sigma) for session hijack, token theft, and impossible-travel/anomalous-OAuth patterns
  2. A precursor-hunting playbook for the credential-to-ransomware window — 50% of ransomware victims had an infostealer or credential exposure event within the prior 95 days (DBIR 2026)
  3. A help-desk verification protocol that survives a cloned voice — because it never relies on recognising one
  4. An IR playbook for identity compromise: token revocation order, session inventory, blast-radius mapping in M365/Okta/Google Workspace
  5. An LLM-assisted enrichment workflow for identity alerts, built on the Day 1 detector stack

The six modules

#ModuleFocus
2.1The identity-first kill chainInfostealer → initial-access broker → operator hand-off (22 seconds, M-Trends 2026); malware-free tradecraft; where detection opportunities survive
2.2Vishing & help-desk impersonationThe M-Trends #2 vector; Scattered-Spider-style MFA-reset fraud; cloned voices (Pindrop: 1 attempt/month → 7/day); Arup as the ceiling case; verification workflows that don't depend on human ears
2.3MFA bypass, session hijack & token theftAiTM phishing kits, OAuth token harvesting, long-lived session abuse; detection engineering on identity-provider telemetry
2.4Valid-account abuse in cloud35% of cloud incidents (CrowdStrike); cloud-conscious intrusions +37%; separating legitimate admin behaviour from adversary use of the same credentials
2.5Precursor hunting: the 95-day windowHunting credential-exposure and infostealer signals before ransomware lands; LLM-assisted correlation of exposure feeds with identity telemetry
2.6IR playbook: identity compromiseOut-of-band verification, token revocation sequencing, session inventory, comms; drafting the incident report with the Day 1 copilot — with evidence-integrity rules

Lab 2

Blue-on-identity: from help-desk call to containment

Students receive replayed identity-provider telemetry (Entra ID / Okta-style logs) covering a compressed attack: an infostealer exposure 60 days earlier, a vishing call transcript against the help desk, an MFA reset, AiTM session capture, and valid-account cloud activity. Tasks: build and tune the session-hijack and token-theft detections (Sigma), hunt the precursor signals backwards, run the identity-compromise IR playbook forward — including a live out-of-band verification drill where one call in three uses a cloned voice. Scoring: time-to-verdict per stage, false-positive rate of the shipped detections, and completeness of the revocation sequence.

The cloned-voice element is deliberately un-detectable by ear — the lab's lesson is that the control is the verification workflow, not human perception.

Key references for Day 2

Version note: the v1 Day 2 spec (deepfake audio/video detection focus) is archived here, including its module deep-dives. Its synthetic-media detection material survives as case-study content inside module 2.2.