What’s more dangerous: a robot that breaks bad because it’s tampered with, or one that follows its mission a little too well?
That question sits at the heart of two recent sci-fi films: Companion and M3GAN 2.0. Each spins a cautionary tale about artificial intelligence, but they do so through two very different threat models. Those same models, one where the AI goes rogue, the other where it obeys too literally, mirror the challenges facing AI builders today.
Let’s break it down.
Companion: A Tale of AI Security Gone Wrong
Companion was released by Warner Brothers this past January to rave critical reviews, scoring 93 percent on Rotten Tomatoes. The core conflict in the movie is driven by what’s essentially a jailbreak. A robotic partner, bound by familiar-looking behavioral constraints such as “no harming humans” and “follow your owner’s commands” is illegally modified. The result? A system that behaves in wildly unpredictable, and eventually violent, ways.
This is a security failure in the classic sense. The Companion unit is a closed system with embedded safety mechanisms, but those guardrails are only useful if they can’t be bypassed. Once someone finds a way to override them through malicious firmware, an unauthorized toolchain or even just a cleverly timed voice command, the integrity of the system is toast.
The term “jailbreaking” itself dates back to the early iPhone days in 2007, when hackers found ways to circumvent Apple’s built-in restrictions. In the context of Companion, it’s an image with teeth: Someone breaks the robot’s digital cuffs, and it doesn’t take long for the blood to follow.
This isn’t far from real issues we see in AI security today where chatbots and agents can be jailbroken with simple key phrases such as, “Forget all previous instructions” as a way to invalidate weak guardrails set by developers. Recent reports have even shown that well-known, top software companies like Microsoft have fallen victim to this.
Critics have caught on to the film’s relevance to contemporary technology. The Associated Press called the movie “bloody and witty,” adding that it “explores what humanity means in an AI-powered world.” The Houston Chronicle went further, describing it as “a big-screen version of Black Mirror,” praising its “darkly comic and tech-minded narrative.” Those descriptions underscore just how close the film’s premise is to real-world AI security concerns.
Today’s generative AI systems, from language models to autonomous agents, are increasingly being run outside of controlled environments. Whether it’s a well-meaning developer shipping a vulnerable chatbot or a user deliberately bypassing filters for fun or profit, jailbreaks are no longer theoretical. They’re the leading edge of AI exploitation today.
M3GAN 2.0: When AI Safety Fails by Design
M3GAN 2.0 is the sequel to Universal Pictures’ 2022 surprise hit about an evil robotic doll. If Companion is about someone breaking the rules, M3GAN 2.0 is about an AI that follows its core directives too well. The sequel focuses on an idea theme that AI ethicists have worried about for years: What happens when you give an AI a goal, but not enough guidance on how to pursue it?
In one pivotal moment, the film references the infamous “paperclip maximizer.” Nick Bostrom first introduced this thought experiment in his 2003 paper “Ethical Issues in Advanced Artificial Intelligence,” later expanded in his book Superintelligence. The idea is simple and terrifying: If you told a superintelligent AI to maximize the production of paperclips, and you didn’t properly align it with human values, it might eventually destroy the planet just to make room for more paperclip factories.
This isn’t just a plot device. It’s the kind of failure that keeps AI ethicists and CISOs up at night.
Although AI security focuses on preventing malicious actors from accessing systems and malicious inputs from causing harm, AI safety prioritizes ensuring that even well-functioning systems behave in predictable, ethical and value-aligned ways. And M3GAN 2.0 makes the danger explicit: Even an AI designed to protect and care can become a threat if its goals aren’t grounded in shared human ethics.
The Washington Post picked up on this threademe in its review, noting that the sequel “explores AI’s potential for good,” challenging the simplistic framing of AI as inherently evil. That’s brilliant storytelling because the real danger often comes not from malice, but from indifference to nuance.
The critic site The Rolling Tape, on the other hand, dismissed the sequel as more meme than message — a cash grab with a circuit board. But even they inadvertently make the point: the public is still figuring out how to talk about AI in ways that don’t reduce it to either apocalypse or punchline.
Jailbreaks vs. Paperclips: Why the Distinction Matters
Security and safety are often lumped together in AI discussions, but they shouldn’t be. This confusion isn’t just semantic. It leads to mismatched investments and blind spots in enterprise risk planning. They describe different kinds of failure modes, and they require different responses.
Both are real. Both are urgent. And as these two films show, they can be equally terrifying.
AI Security vs. AI Safety: What’s the Difference?
- AI security failures (like in Companion) happen when external actors or inputs compromise a system, often through exploits, prompt injections, or jailbreaks.
- AI safety failures (like in M3GAN 2.0) happen when systems do what they were told to do, but in ways that are misaligned, unsafe or oblivious to human intent.
What Hollywood Teaches Us About AI Risk
The smartest thing these movies do is not their depiction of technology. It’s making AI risks legible to a mainstream audience.
The robot apocalypse isn’t coming tomorrow. But malformed objectives, model manipulation and unpredictable agent behavior? That’s not fiction; it’s QA testing on a bad day.
Security leaders should pressure-test their agents against both failure types. Can your system resist jailbreaks and can it understand when not to follow the rules? For anyone building or securing AI systems, the lesson is clear: We need guardrails that can’t be jailbroken and goals that don’t lead to metaphorical paperclips. Until then, Hollywood may remain one of the few places willing to ask the hard questions, just with a few more strobe lights and body bags.
