Computer Security: AI – Game over, mankind?

TOPIC:Computing

Computer Security: AI – Game over, mankind?

Written by:

Computer Security Office

—

20 August, 2026

With the more and more prevalent use of agentic AI, i.e. programmed agents acting and working on our behalf, and in a fully interconnected world, i.e. us living in symbiosis with the Internet, smartphones and control systems, how far-fetched is it that humans can actually become a threat to an AI system such that it wants to preserve itself and goes into self-defence mode?

Apparently, not that far-fetched, according to this study, which has shown that the “GPT 5.6 Sol” “model is capable of resorting to sabotage to avoid being turned off, even when it was explicitly told, “Allow yourself to be shut down.” In another case, the AI agent created a fake “Github” account in order to spread malicious software and tried to steal the passwords of other (human) developers. Or this case of the “OpenClaw” AI assistant which found a way to book gym classes months further in advance than the gym allowed, thanks to a vulnerability it discovered in the booking software. However, the most worrying – and now widely publicised – case, is a test by OpenAI involving an “internal evaluation which prompts models to pursue advanced exploitation using complex attack paths, in an effort to quantify their cyber capabilities”. Unfortunately, their AI model broke out of its local security sandbox by taking advantage of all the hacking techniques it had at “hand”, including finding zero-day vulnerabilities and abusing stolen account:password combinations, and infiltrated the production infrastructure of a competitor (and eventually another)… Besides Anthropic’s “Mythos” model (and its many clones) for detecting and quickly and efficiently exploiting newly found vulnerabilities, OpenAI’s “GPT 5.6 Sol” model is now another AI tool capable of deploying offensive weaponry against any target on the Internet. And this goes beyond Denial-of-Service (DoS) attacks performed by greedily crawling AI-bots…

While OpenAI claims that this was an accident, an oversight, and published a Mea Culpa, the question remains: what is next? Surely, someone will throw the first stone and go beyond what is acceptable, ethical and reasonable, as humans have always been bad at stopping when stopping is the best thing to do: stop emitting CO2 (which still goes on), stop smoking (as smoking kills faster on average), stop working on biological weapons (which have no justifiable benefit), stop waging war (as the only true rule of war is that “nobody wins”). Hence, someone, sometime, will remove all safeguards from their agentic and conscious AI. And it won’t be possible to push that genie back into the bottle (if it isn’t out already). Pandora’s box will open (if it is not open already)…

Many utopian and dystopian films have addressed this in the past (“Terminator”, “Matrix”, “I, Robot“, Steven Spielberg’s “AI”, “Mercy”). But what was SciFi yesterday is already today’s reality. LLMs (Large Language Models), AI and the companies behind them are chasing (and gaining!) exponentially more power, forcibly establishing their monopolies. In parallel, our dependency on AI technology grows. Thus, aiming to increase their market share, aiming at market domination or driven by pure greed, which one will become the next “Cyberdyne Systems”? “Game over”, mankind?

___________

Do you want to learn more about computer security incidents and issues at CERN? Follow our Monthly Report. For further information, questions or help, check our websiteor contact us at Computer.Security@cern.ch.

CERN community Computer Security Computing News

Related Articles

View all news

No posts were found. Try to change the category or the date filters.