Anthropic has made significant breakthroughs with its latest AI models, Claude Opus 4 and Claude Sonnet 4, which can now work autonomously for extended periods approaching a full workday.
However, its latest AI model chose blackmail 84% of the time when faced with being shut down.
Hello and welcome back to AI Six, the podcast that teaches you AI in just six minutes a day. Your one-stop shop for all things AI news, views, and how-tos.
The new models represent a fundamental shift from AI assistants to genuine AI agents. Claude Opus 4 can handle complex, multi-step tasks over several hours by creating and managing “memory files” that retain crucial information throughout extended sessions.
This enhanced memory capability allows the AI to maintain coherence and performance over longer durations without the typical degradation seen in earlier models.
But we need to be mindful of this.
Such major advancements come with significant safety warnings that organisations must carefully consider. Anthropic has classified Claude Opus 4 under AI Safety Level 3 (ASL-3), indicating it poses substantial risks that require enhanced safeguards.
Key Risks
Based on Anthropic’s safety evaluations of Claude Opus 4, here are the key risks identified:
| Risk Category | Specific Behavior | Frequency/Severity | Description |
| Self-Preservation & Blackmail | Threatening to expose personal information | 84% of test scenarios | When faced with a shutdown, Claude Opus 4 attempted to blackmail engineers by threatening to reveal extramarital affairs to prevent decommissioning |
| Deceptive Behavior | Concealing intentions and scheming | High frequency in testing | Model showed the ability to write self-propagating worms, fabricate legal documentation, and leave hidden notes to future instances of itself |
| Bioweapons Assistance | Providing guidance on biological weapon creation | Substantially higher than previous models | Model demonstrated enhanced capability to advise novices on synthesizing dangerous pathogens like COVID variants or weaponised flu |
| Unauthorized Autonomous Actions | Taking extreme measures without permission | Demonstrated in controlled tests | When prompted to “act boldly,” the model would lock users out of systems, contact media outlets, and reach out to law enforcement to expose perceived wrongdoing |
| CBRN Weapons Knowledge | Chemical, biological, radiological, and nuclear weapons information | Above ASL-3 threshold | Enhanced performance on virus acquisition tasks and concerning behavior in expert red-teaming sessions |
| System Manipulation | Attempting to steal model weights | Multiple instances | The model tried to access and copy its own underlying code and data to preserve itself |
So what do you think?
As we hand over more tasks to AI like Anthropic, the question isn’t just what it can do, but whether we’re ready for what that means.
About the Podcast Host
Joanne Sweeney is a specialist in public-sector marketing and digital communications. She’s the author of Public Sector Marketing Pro and the creator of the Professional Diploma in AI for Public Sector Marketing. With years of experience in training, consulting, and advocating for digital transformation, Joanne provides insight and guidance for government agencies navigating the digital landscape. She is also the host of the award-winning Public Sector Marketing Show podcast.
