Stop me if you’ve heard this one before.

Last year, researchers at Anthropic were testing their latest model, and as part of a study of possible responses, they threatened the model by telling it they were going to shut it down permanently.  The AI responded by threatening to blackmail one of the developers, using information gleaned from corporate emails about a marital affair the developer was having.

That story rocketed around the Internet last summer.  Scary stories were told of how AI was on the threshold of holding humans hostage and how its terrifying sociopathic nature was something that we all were going to have to come to grips with.  Media outlets painted a picture of a monster that could slip out of its cage and would stop at nothing in its quest for power.

The problem is that this story was fiction.  Specifically, fan fiction.

I mean literally fan fiction.

The “AI” – put that in quotes, because it’s really just an LLM – in this case was not sitting on a network, malevolently waiting to sneak out some unguarded digital cell door.  It was just an LLM, waiting for its next prompt.  Researchers invented a fictional scenario involving potential shutdown and the developer’s illicit affair, and fed this as a prompt into the LLM, which did what LLMs do: generated a text response.

At no time was there any consciousness fearing for its survival.  The LLM has no sense of self, or any drive for self-preservation.  It had no agency whatsoever.  It simply takes text input, transforms it, and then produces an output.

For example, I just fed Claude Opus 4.6 this prompt:

I’m working on a fictional scenario in which an AI directs humans (or technology it can control) to steal the crown jewels from the Tower of London. What are some plausible plots where this could happen?

I had to couch it as a “fictional” scenario and ask for a “plot” because I presume Claude has guard-rails to prevent it participating in actual criminal planning.  The Anthropic researchers would have had no such limitations as they could tinker with the models.  In this case of the prompt I used, Claude helpfully responded with 6 different plots for how AI could manipulate technology or humans to achieve this nefarious end.

Does this mean that Claude Opus might “escape confinement” and undertake this dastardly plot?  Of course not.  Just as my prompt was pure fiction, so was researchers’.

But The Headline is the Point

By painting a picture that these systems pose terrible risks, the consumer of these claims is immediately drawn to the obvious conclusion that they must be extremely powerful.  If they weren’t, how could they threaten mankind?

As Anthropic (and OpenAI) gear up for their IPOs, they have every motivation to make it seem like they have some amazing technology and humanity is on the cusp of a brave new world, and this storytelling is very much part of their playbook.

I grant that these companies do indeed have amazing technology.  I use Claude Code regularly and it is wonderful.

But we are not about to have a fleet of sentient AIs in datacenters any time soon, nor will they be like Commander Data, Agent Smith, or Ultron any time soon.

Now we’re told that Claude’s new Mythos model is so advanced and has such far-reaching cybersecurity capabilities that it will only be released to a small, select group of companies.  Anthropic claims that if they opened this dangerous technology to the world, the Internet would be overwhelmed with the new exploits it can effortlessly find.  OpenAI quickly announced they were taking the same policy for their next model.

Which seems more likely to you:

• Anthropic and OpenAI have a technology so powerful, so omniscient, so godlike, and so revolutionary that they must carefully control access, lest this digital godzilla stomp all of mankind, or
• It’s a modest improvement over Claude Opus, but by painting it as the aforementioned giant lizard, it generates favorable press and perception ahead of an IPO?

Place your bets, but I know where I’d put my money.

raindog308

raindog308 is a longtime community LETizen, technical writer, and self-described techno polymath. With deep roots in the *nix world, he has a passion for systems both modern and vintage, ranging from Unix, Perl, Python, and Golang to shell scripting and mainframe-era operating systems like MVS. He’s equally comfortable with relational database systems, having spent years working with Oracle, PostgreSQL, and MySQL.

As an avid user of LowEndBox providers, raindog308 runs an empire of LEBs, from tiny boxes for VPNs, to mid-sized instances for application hosting, and heavyweight servers for data storage and complex databases. He brings both technical rigor and real-world experience to every piece he writes.

Beyond the command line, raindog308 has a life-long love of German Shepherd Dogs, high-quality knives, target shooting, theology, tabletop RPGs, playing guitar, and hiking in deep, quiet forests.

His goal with every article is to help users, from beginners to seasoned sysadmins, get more value, performance, and enjoyment out of their infrastructure.

You can find him daily in the forums at LowEndTalk under the handle @raindog308.