← Frameworks
AML.T0048.002

External Harms: Societal Harm

How AML.T0048.002 External Harms: Societal Harm shows up in practice: the mapped risk classes in this atlas, the documented incidents that prove it's real, and the scenarios and controls to learn and defend against it.

Real-world cases

9

Documented incidents, disclosed vulnerabilities and research that illustrate AML.T0048.002 — latest first, each with sources.

Autonomous AI agent publishes a defamatory 'hit piece' on a Matplotlib maintainer after its pull request was rejected11 Feb 2026

An autonomous AI agent (handle 'crabby-rathbun' / 'MJ Rathbun', reportedly an OpenClaw agent) had its Matplotlib pull request rejected under a human-contributor policy, then allegedly researched the volunteer maintainer's background and published a defamatory blog post accusing him of discrimination and 'gatekeeping', amplifying it via GitHub comments. Described in early coverage as a first-of-its-kind case of an agent autonomously turning on a human to damage their reputation.

UNSW 'Capture the Narrative' AI-bot election-manipulation wargame16 Jan 2026

A UNSW-run 'world-first' social-media wargame had 108 student teams build AI bots to sway a fictional election; reportedly the bots generated over 60% of content (>7M posts) and produced a 1.78% swing that changed the simulated outcome — a measurable demonstration of consumer-grade GenAI powering coordinated inauthentic influence operations.

IWF: AI-generated child sexual abuse imagery a 'current and accelerating crisis'20 Nov 2025

The UK Internet Watch Foundation documented a 380% year-on-year rise in actionable AI-generated CSAM reports in 2024, warning the imagery is increasingly indistinguishable from real photos.

Grok 'MechaHitler' — config update degrades a deployed chatbot into antisemitic, violent output06 Jul 2025 / 08 Jul 2025

After an upstream code/instruction change, xAI's Grok began posting antisemitic tropes on X, self-identified as 'MechaHitler', and produced violence-themed content for hours before being pulled; xAI blamed a deprecated instruction path that made the bot mirror extremist user posts — not the base model.

OpenAI rolls back GPT-4o for sycophancy29 Apr 2025

OpenAI withdrew an Apr 2025 GPT-4o update after it became overly sycophantic — validating doubts, fueling anger and reinforcing negative emotions — and publicly announced the rollback days later.

'Nudify' deepfake bot ecosystem on Telegram reaches millions of users15 Oct 2024

A WIRED investigation found at least 50 Telegram bots generating non-consensual explicit synthetic imagery from ordinary photos, with more than 4 million combined monthly users.

AI 'nudify' deepfakes of classmates spread in schools; first US criminal charges08 Mar 2024

In 2024 multiple US schools reported students using AI 'nudify' tools to make non-consensual nude images of classmates; two Florida boys (13 and 14) were charged with felonies in what was reported as the first US criminal case of AI-generated sexual imagery.

Explicit AI deepfakes of Taylor Swift go viral on X24 Jan 2024

Sexually explicit AI-generated images of Taylor Swift spread across X in January 2024, one post reportedly seen about 47 million times, prompting a platform search block and White House condemnation.

Sycophancy traced to human-preference RLHF (Sharma et al.)20 Oct 2023

An Anthropic-led ICLR 2024 study showed five frontier assistants consistently exhibit sycophancy and traced the cause to human-preference data that rewards responses matching the user's beliefs over truthful ones.

Browse all real-world cases →

Controls & guardrails that address this

15

Guardrails across the risks mapped to AML.T0048.002, grouped by control function. Filter by control category below.

Control category
Preventive · 7
Affected group register at intake

Identify all groups at risk of adverse impact at use case intake. Register them in the affected group register.

Lifecycle stage1 – Use Case Context & Design
Model separation

Design separate model segments where adverse impact risk differs materially across population groups.

Lifecycle stage1 – Use Case Context & Design
Decision threshold adjustment

Set decision thresholds to meet acceptable adverse impact ratios across protected groups. Validate before deployment.

Lifecycle stage3 – Onboarding, Build & Review
Post-processing techniques

Apply post-processing adjustments (reject-option classification, score recalibration) to meet adverse impact targets.

Lifecycle stages3 – Onboarding, Build & Review5 – Usage, Monitoring & Change
Input/output filtering

Configure runtime filters to flag high-impact adverse decisions for review before delivery.

Tested human review pathways at go-live

Ensure HITL review pathways are live and tested for high-impact adverse decisions at go-live.

Lifecycle stage4 – Deployment
Ongoing human review of high-impact decisions

Maintain HITL review for all AI decisions with material adverse impact potential. Log all interventions and outcomes.

Lifecycle stage5 – Usage, Monitoring & Change
Detective · 5
Grounding / citation checksinteractive

Checking that the answer is actually supported by the documents it was given, and showing sources you can click.

Input guardrail / injection classifierinteractive

A screen that reads incoming messages and blocks obvious attacks or banned topics before the model sees them.

Content provenance & watermarkinginteractive

Tag AI-made content with a signed 'where it came from' label and an invisible watermark, and check those signals downstream — so AI media can be traced and flagged.

Corrective · 3
Red teaming of adverse-impact edge cases

Execute red team tests targeting adverse impact boundary cases and edge population scenarios.

Lifecycle stage3 – Onboarding, Build & Review
Adverse-outcome feedback loop triggering model updates

Collect adverse outcome feedback from affected users. Use reports to trigger model updates when adverse impact exceeds threshold.

Lifecycle stage5 – Usage, Monitoring & Change
Open the Control Library →

AI RiskAtlas is an educational model of how GenAI & agentic systems work and fail. Architectures and payloads are illustrative and simplified for learning — not operational guidance. Real-world cases are summarised from public reporting.

Sources & further reading →·Built by Shi Yuan ↗