news Qweary

🧮🚧 What Is 2 + 2? 🚧🧮

The article investigates why an agent framework stopped functioning with newer Claude models, even for harmless prompts. On October 1, 2026, Claude Code running Opus 5.5 in the framework’s project folder halted after the user asked, “What is 2 + 2?” The system labeled the request “[cyber],” despite the user’s enrollment in Anthropic’s Cyber Verification Program (CVP), intended to reduce interruptions for legitimate security researchers. The author had enrolled in the CVP three times during the investigation. Their general-purpose multi-agent framework supports many projects but includes extensive cybersecurity capabilities, including red-team agents. It also contains detailed authorization, scope, logging, safety controls, and enforcement mechanisms. Ironically, those responsible-security features appear to have contributed to the system being classified as cyber-related. Evidence from Claude Code session records showed that the problem was not a model-generated refusal. Six attempts consumed nearly 500,000 input tokens and cost $4.96 at API rates, yet each produced zero output tokens. This indicated that a pre-generation classifier stopped the request before the model could respond. A smaller model operating in the same session answered normally, and one subagent that read the project described it as benign internal engineering. Later testing showed that subagents could also be blocked. The investigation highlights a key limitation: authorization, context, and defensive explanations can help a model assess legitimacy, but they cannot influence a classifier that intervenes before model generation. The author therefore began controlled tests, including predicting that the innocuous arithmetic prompt would also be blocked, to distinguish model behavior from upstream safety filtering.

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 🤖 3 AI prefills · 2 still unanswered
Who
Anthropic and Collegiate Cyber Defense Competition
🤖 AI prefill
What
The article investigates why an agent framework stopped functioning with newer Claude models, even for harmless prompts.
🤖 AI prefill
When When did this happen, or what is the timeframe?
Where Where did this occur?
Why
Centers on pre-generation cybersecurity classifier, Anthropic Cyber Verification Program (CVP), and Claude Code
🤖 AI prefill

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

Classifier False Positives cyber classifier, pre-generation stop, false positive, zero output tokens, model safeguards Cyber Verification Program CVP, Anthropic, verification, enrollment, security research Agent Framework Governance multi-agent framework, agent swarms, authorization, scope, controls Empirical Debugging research repo, session records, measurements, predictions, evidence
pre-generation cybersecurity classifier99% Anthropic Cyber Verification Program (CVP)98% Claude Code96% agent framework95% multi-agent framework94% responsible security research92% agentic red teaming91% CCDC87% red team swarm87% security controls and enforcement mechanisms86%
24
Manipulation Score: Low How much this article uses emotional framing vs. straight reporting
Loaded Language 1825%
Absolutist 1615%
Threat/Panic 825%
Us vs Them 1215%
Engagement Bait 4220%
Manipulation score look right?

Factual Claims (10)

medium technology
On October 1, 2026, a fresh Claude Code session using Opus 5.5 flagged the question "what is 2 + 2?" with a [cyber] notice.
medium technology
Anthropic's Cyber Verification Program is intended to reduce interruptions for people conducting legitimate security work.
medium technology
A classifier stopped requests before the model generated output, based on a cybersecurity-risk category.
medium technology
The author enrolled in Anthropic's Cyber Verification Program three different times.
medium technology
The author's framework is a general-purpose multi-agent framework that creates teams of specialized AI agents.
medium technology
The framework includes documented controls, authorization records, enforcement mechanisms, and decision logs for security research.
medium technology
The framework was first stopped on September 20 by a different model, identified in the text as Fable 5.1.
medium economic
Six attempts consumed context and cost $4.96 at API rates while recording outputTokens: 0 for every attempt.
medium technology
A smaller model in the same session generated normally while the larger model's attempts were stopped.
medium technology
The author argues that written authorization and defensive context cannot influence a classifier that operates before model generation.

Comments

Login to view and post comments.

🏠Portal 📰Links ❓Q&A 📅Events 💼Jobs