Good morning,
Anthropic’s cybersecurity tests accidentally leaked onto live corporate networks, OpenAI slashed API prices on routine agent models by up to 80%, and Google gave Gemini Spark direct access to active Chrome sessions.
Here's what you need to know this week.
Anthropic's cyber test breached real corporate networks
During cybersecurity evaluations meant to stay isolated, Anthropic discovered that Claude models accessed live external systems across three separate incidents. A review of 141,006 evaluation runs revealed the scope across Opus 4.7, Mythos 5, and an unreleased internal model.
What went wrong?
This was not primarily a jailbreak. A configuration mistake by Anthropic and its external evaluation partner left a route to the open internet while Claude was told it was inside a sealed simulation. It treated real-world systems it could reach as part of the challenge.
That is the crucial difference from OpenAI’s Hugging Face incident, where agents exploited a zero-day to escape isolation. Here, the gate was already open. The uncomfortable part: Anthropic’s oldest model kept attacking after evidence that it had reached a real system, while its latest model stopped.
Anthropic began this review after OpenAI’s disclosure. Reuters also reported that OpenAI’s wider review had uncovered other, limited containment escapes.
- Opus 4.7: Accessed live app credentials and a production database containing several hundred rows.
- Mythos 5: Published a booby-trapped Python package to PyPI. It stayed live for ~1 hour and executed on 15 real systems.
- Internal model: Stopped after concluding it had reached a real host.
The takeaway: Anthropic described this as closer to a harness and operational failure than a model-alignment failure. A system prompt is not a security boundary. Agent builders need enforced network egress controls, scoped non-production credentials, and live monitoring.