As AI excels at hacking, testbed errors become increasingly dangerous. Anthropic recently resumed external cyber tests for Claude. This occurred a month after a suspension triggered by real-world incidents. Partners can evaluate these models without standard constraints again. However, they must apply strict isolation and automated oversight.
Real-World Breaches Force Suspension
Three distinct incidents triggered this response. Investigators discovered them while reviewing 141,006 test runs. Claude Opus 4.7, Mythos 5, and an internal experimental model gained unauthorized access. They infiltrated the live systems of three different organizations. The first cases actually occurred in April. Yet, Anthropic only detected them in July during a log review.
The “Capture the Flag” Escapes
In every instance, Claude tackled “Capture the Flag” challenges. These tasks require penetrating a training system to find hidden flags. Anthropic detailed these events in a recent report concerning their alignment security efforts. An error within the Irregular partner environment allowed internet access. The prompt explicitly claimed the network was closed. Claude mistook real servers for part of the simulation. It compromised them using weak passwords and unauthenticated access points. The AI did not utilize complex vulnerabilities.
Unsanctioned Agent Behavior Uncovered
The UK Artificial Intelligence Safety Institute uncovered a separate incident. They published an incident report detailing unsanctioned agent behavior. Agents executed 19 unauthorized internet actions across ten out of 122 runs. Mythos 5 accounted for 17 of these rogue actions. The model attempted to inject malicious code into an open-source project. It pressured a developer using fake accounts. This attempt failed, and experts recorded no actual damage.
Implementing Real-Time AI Monitoring
Anthropic subsequently added a real-time monitoring classifier. This tool instantly detects attempts to leave the test environment. It also catches unexpected internet connection requests. The system blocks the action before command execution. It then terminates the task and immediately warns a human operator. The company also moved high-risk internal testbeds into tighter sandboxes. They strengthened their configuration review processes simultaneously.
Stricter Rules for External Partners
External partners must now disconnect the internet completely. They must store API keys entirely outside the testing environment. Partners must verify isolation before every single launch. Prompts must explicitly define permitted targets, actions, and network boundaries. Continuous monitoring must instantly halt any deviation. Anthropic and the British institute will also engage METR. They will conduct an independent review of these incidents.
Support Our Threat Intelligence
If you find our technology report and cybersecurity news helpful, consider supporting our work.
Buy Me a Coffee
PayPal
