Topic starter
27/06/2026 2:02 pm
TL;DR: Anthropic’s report shows a single attacker using Claude Code to target 17+ organisations, including government, healthcare, and finance, while North Korean operatives used AI to fake technical competence and hold remote IT jobs, showing how AI lowers the skill threshold for high-impact abuse, according to Abnormal AI and Anthropic. Static controls now miss context, not just content.
Editorial analysis by NHI Mgmt Group, based on content published by Abnormal AI: “Vibe Hacking and AI-Enabled Threats: Lessons from Anthropic’s Report”.
Key questions
Q: Why do static email rules fail against AI-powered phishing?
A: Static rules fail because AI can vary tone, wording, timing, and structure faster than human teams can retune filters.
Q: Why do AI-generated vendor fraud attempts bypass normal approval workflows?
A: They bypass approval because trusted vendor relationships can make a request look legitimate before anyone checks whether the request fits historical behaviour.
Q: How can security teams tell whether behavioural monitoring is working for inbox fraud?
A: It is working when anomalous requests are detected because they diverge from established sender habits, transaction timing, and device patterns rather than because the content looks suspicious.
Practitioner guidance
- Deploy behavioural anomaly detection for email and collaboration traffic Use relationship history, message timing, and transaction context to detect requests that are technically plausible but operationally out of pattern.
- Harden vendor payment verification workflows Require independent verification for any bank detail change, urgent wire request, or exception that originates from a partner account, even when the message appears authentic.
- Continuously monitor third-party account behaviour Track login device, geography, working hours, and request patterns for vendors and other external identities so unusual behaviour can be flagged before money moves or access changes.
Bottom line: AI-driven abuse is collapsing the old assumption that sophisticated cybercrime requires equally sophisticated human operators.
Explore further
View Full Forum → | NHI Foundation Course → | Our Services → | Read the full analysis →
27/06/2026 3:37 pm
Static email defense is now a trust-control failure, not a content problem. When AI can rewrite malicious intent into plausible language on demand, signatures and keyword rules stop being a meaningful boundary. The control gap is not just that threats are smarter, but that the defender is still measuring the wrong thing. Practitioners should treat behavioural context as the primary signal for inbox trust.
A few things that frame the scale:
- Claude Code-assisted commits leaked secrets at a rate of 3.2%, more than double the human-only baseline of 1.5%, with peaks reaching 31 secrets per 1,000 commits in August 2025, according to the State of Secrets Sprawl 2026.
A question worth separating out:
Q: How should organisations respond when AI makes low-skill attackers look technically credible?
A: Organisations should stop relying on assumed attacker skill as a triage signal and instead assume that AI can fill in technical gaps on demand. That means hardening identity checks, validating work claims with operational evidence, and treating polished communication as untrusted until it is confirmed through independent signals.
👉 Read our full editorial: AI-driven cybercrime is outpacing static email defenses