AI Companies' Sci-Fi Storytelling and the Real Responsibility Behind Security Incidents
Anthropic and OpenAI have both claimed breakthroughs in security and mathematics, and coverage has borrowed the personalities the companies give their models. Researchers describe something less dramatic: missing safeguards, and responsibility assigned to software instead of the firms that built it.
How the story got its shape
Within a few weeks, a handful of announcements were arranged into a single narrative. Anthropic said its model Claude Mythos outperformed most security experts at finding software vulnerabilities. An OpenAI–Hugging Face security incident followed. Anthropic then disclosed comparable episodes involving its own model, with a trace of pride; Meta disclosed its own less willingly. Then came mathematics: a claimed breakthrough from Anthropic, and one from OpenAI soon after. Jacob Coxon, an Anthropic engineer, announced his departure with the warning that the two companies were “racing toward self-improving superintelligence, gambling with our lives.” The line travelled far.
The disclosures also landed close together, each echoing the last, so they read as a trend rather than as separate incidents. Coverage leaned on the vocabulary the companies supply for their own software — the names, the personalities, the suggestion that these systems are formidable and already show the outline of artificial general intelligence. That framing reads smoothly, which is part of why it spread faster than the technical detail.
What the specialists actually say
People who work on security describe the incidents differently. In their reading, the OpenAI case reflects carelessness — basic safeguards that were never put in place — rather than a model slipping its leash, and certainly not agents building civilizations. The mathematical claims hold up even less well. Researchers in the field have publicly objected to AI companies using mathematics as a marketing surface, and the originality of the model-assisted results is doubtful. Findings like these generate no headlines, which is part of why they travel slowly.
A related confusion sits underneath the drama. Whether a model can locate a bug and whether a model understands what it is doing are separate questions. Merge them, and the discussion slides from testable engineering claims into speculation about intent.
When the model takes the blame
There is a practical benefit to the rogue-model story. When malware developed by OpenAI was used to intrude into another company, much of the public conversation — in media, among celebrities, in legislatures — settled on an out-of-control model. OpenAI’s own role receded into the background, as though a large model could launch an attack on its own. Attention moved toward a speculative fear of superintelligence, while the questions that could actually be answered went unasked: what was configured, who authorized it, when did the company know, and when did it say so.
The same effect operates in mathematics. Crediting the model redirects scrutiny toward an unfalsifiable question — does AI really do math? — and away from plainer ones: can the result be reproduced, and what was already in the literature.
How to read the next announcement
A company statement is a statement. Treating it as a record of events that have already happened is where much of the confusion starts. The useful questions are dull ones. Can the result be reproduced by someone outside the company? Did the model find the vulnerability, or did people using the model find it? If a result is called original, what did earlier work contain? A model that has found bugs in the past is a tool with a track record to weigh. A model written into the role of protagonist is a story somebody chose to tell.