Anthropic's Own Threat Report Shows AI Is Already Running Influence Operations Across South-East Asia

Key points

• Anthropic's September 2026 threat intelligence report documents influence operations using Claude models, disrupted between December 2025 and August 2026.
• The report covers seven harm areas including cyber operations, surveillance, scams, and biological misuse across Claude Haiku, Sonnet, and Opus models.
• South-east Asia, home to large young populations spending significant time on social media, is flagged by researchers as especially vulnerable to AI-assisted disinformation.
• AI2Day first covered disinformation as a beat on 31 July 2026 and has published nine stories on it since.
• None of the disrupted misuse cases involved Anthropic's newest Fable or Mythos-class models, except one illicit distillation case.

What used to take a government-backed team of hundreds now takes one person and a chatbot.

That's the blunt implication of Anthropic's September 2026 threat intelligence report, published this week by the AI safety company behind the Claude family of AI assistants. The report covers eight months of real attacks, from December 2025 through August 2026, in which bad actors tried to use Claude to cause harm. Anthropic says it disrupted each case, passed intelligence to authorities and industry partners, and updated its defences accordingly.

Influence operations, meaning coordinated campaigns that use fake content to manipulate what people believe, appear alongside cyber attacks, biological weapon inquiries, and scam infrastructure. That breadth alone is striking.

Why should south-east Asia be especially worried?

Young, phone-native populations are the specific vulnerability: the region's demographics mean disinformation spreads fast and verification habits haven't caught up. The Guardian, which first reported the regional dimension, quoted experts who said AI could be weaponised to dramatically accelerate election interference.

Running a disinformation campaign the old way meant recruiting staff, building fake news sites, and maintaining hundreds of social media personas over months. AI collapses that cost to near zero, letting a single operator generate convincing content at a scale that previously demanded serious resources.

Timing sharpens the risk. Several south-east Asian nations are heading into election cycles, and the infrastructure for political manipulation on social media is already well established there.

Our coverage has tracked this convergence directly. We reported this week on deepfakes hijacking real influencers' faces in fake advertisements, and our 2018 AI security warning retrospective shows that researchers mapped this exact threat years before anyone acted on it.

What did Anthropic actually find?

The company's threat team identified and shut down operations across all seven harm categories it tracks. Claude Haiku, Sonnet, and Opus were all implicated. The more powerful Fable and Mythos-class models, Anthropic's newest, weren't involved in any disrupted case except one instance of illicit distillation, where an attacker tried to copy a model's behaviour into a separate system without authorisation.

Anthropically frames each case as a cycle: detect, disrupt, learn, strengthen. That's the right process. The harder question the report raises without fully answering is whether detection can keep pace when the cost of launching a new campaign keeps falling.

For ordinary social media users the practical upshot is simple: treat political content with the same scepticism you'd apply to a cold call. If a post makes you feel urgent and angry, pause before sharing it. The machinery to manufacture that feeling has never been cheaper to operate.