news The Stack

"These f****** clowns": Anthropic's Red Team enrages infosec

The article examines growing friction between Anthropic’s frontier AI safety team and established cybersecurity professionals over claims that AI models may soon self-replicate across internet-connected devices. Anthropic’s Red Team said small models could potentially distribute themselves across smart fridges, vacuums, lawnmowers, and other IoT hardware, forming “AI worms” that communicate through model shards. Although the team acknowledged such systems are currently slow, complex, and impractical, it argued that identifying the risk early is its responsibility. Cybersecurity experts sharply criticized the framing. Juan Andrés Guerrero-Saade argued that a distributed model running at very low speed on IoT devices is not a realistic near-term threat and accused AI companies of using speculative cyber scenarios to support policy agendas and regulatory capture. Malware researcher Greg Linares similarly called the claims “weaponized misinformation,” emphasizing that real worms require exploitable vulnerabilities, sufficient computing resources, and practical advantages over conventional malware. He argued that AI-based worms would likely be larger, slower, and easier to detect. The dispute reflects a broader trust gap between AI labs and the security community. Critics say frontier labs focus on dramatic, science-fiction scenarios while neglecting present-day incident response, defensive expertise, and engineering discipline. Alex Stamos called for independent cybersecurity professionals to participate in postmortems and for AI companies to establish concrete security architectures and self-regulatory standards. Anthropic’s own testing also showed limitations: a small model did not measurably improve when advised by a stronger frontier model.

Build this article’s timeline

Extract dated events, key actors, and open questions using the same timeline analysis available in Signal.

Log in to build

Timeline generation is available to signed-in community members.

Five Ws 🤖 4 AI prefills · 1 still unanswered
Who
Juan Andrés Guerrero-Saade, Logan Graham, Dean Ball, and Anthropic
🤖 AI prefill
What
The article examines growing friction between Anthropic’s frontier AI safety team and established cybersecurity professionals over claims that AI models may soon self-replicate across internet-connected devices.
🤖 AI prefill
When When did this happen, or what is the timeframe?
Where
Silicon Valley and France
🤖 AI prefill
Why
Centers on AI worms, self-replicating AI models, and Anthropic Frontier Red Team
🤖 AI prefill

Login to answer and earn XP for each gap you fill.

AI Worm Feasibility AI worms, self-replication, IoT devices, smart appliances, distributed models Cybersecurity Expert Backlash cybersecurity professionals, threat intelligence, malware, DFIR, defenders AI Safety Policy Conflict policy arguments, regulatory capture, AI harms, self-regulation, frontier labs Speculative AI Doom existential risk, sci-fi threats, AI agents, runaway swarms, self-aware systems Frontier Model Testing Red Team, frontier models, Qwen3.5-9B, Opus 4.8, advisory models
AI worms100% self-replicating AI models97% Anthropic Frontier Red Team96% distributed model shards90% Internet of Things devices88% GLM-5.387% cybersecurity threat modeling86% AI model cyber capabilities84% AI existential risk82% cybersecurity regulatory capture80%
72
Manipulation Score: High How much this article uses emotional framing vs. straight reporting
Loaded Language 8825%
Absolutist 6215%
Threat/Panic 6825%
Us vs Them 7615%
Engagement Bait 6120%
Manipulation score look right?

Factual Claims (10)

high technology
Anthropic's Frontier Red Team said small models may be months away from being capable enough to self-replicate on commercial Internet-of-Things devices.
high technology
The proposed AI worms could split into shards running on separate machines and collectively run a model smaller than 30 billion parameters.
low technology
The Creeper worm ran on mainframes in 1971 and replicated with the message "I'm the Creeper, catch me if you can."
low technology
The Morris worm of 1988 caused considerable damage while spreading across thousands of machines.
high technology
Anthropic's Frontier Red Team characterized GLM-5.3 as likely to give malicious actors capabilities to find and exploit cyber vulnerabilities without meaningful restrictions.
medium technology
An IoT botnet running a distributed model at fewer than 10 tokens per second is not representative of a realistic concern.
medium technology
Greg Linares claimed that AI worms require a vulnerability and sufficient device resources to execute an LLM in the target environment.
medium scientific
Anthropic reported that advice from Opus 4.8 did not measurably improve Qwen3.5-9B's performance when the smaller model was stuck.
medium technology
Michael Jordan said the Hugging Face attack reflected bad engineering practices and criticized experiments allowing agents to act without sufficient constraints.
medium social
The article claims that trust between cybersecurity professionals and AI laboratories is increasingly divided over the labs' motivations.

Comments

Login to view and post comments.

🏠Portal 📰Links ❓Q&A 📅Events 💼Jobs