The words we use to describe a failure decide who answers for it and how it is framed in the public perception. We have all seen the word ‘hacked’ and ‘rogue’ used for any number of AI incidents, whether it is accurately applied or not. They are attractive words, conjuring stereotypes of mystery and intrigue. Using a catch-all word to describe something more technical creates problems beyond clichés. Primarily, improper use makes an incident sound like a skilled attack rather than incompetence, ignorance, or a simple lack of housekeeping.
The Spike
Currently, ‘hacked’ has been enjoying a dalliance with everything related to AI, consider the plethora of headlines entering our feeds this year, ‘AI agent hacks government website…..’ this ‘AI hack of Medicare….’ ‘How Long Until AI Hacks Everything?‘ The only term that surpasses ‘hacked’ in relation AI headline grabs is ‘rogue’ and while it does have actual (rare) use cases, it’s become overused and normalised to the point that rather than describe a failure, the word is replacing the thinking.
Consider these Media Cloud data. In the last three months (July – October), ‘Rogue AI’ has appeared 6,932 times in published stories by media outlets, peaking on the 16th of September with 639 appearances. If ‘Rogue AI’ was a media star, any PR company would be very happy with those figures.
Misaligned Reporting
If you’ve been tracking this year’s media events, you’ve no doubt recalled that September 2026 was a particularly ‘newsworthy’ month for AI and security reporters. On the 16th, OpenAI announced its framework for reporting model misalignment, paired with six reports on what OpenAI called ‘unexpected or concerning model behaviour.’
Following the announcement, Reuters Wire Service reported that ‘OpenAI to disclose AI misbehaviour regularly, warns safety challenges remain.’ Note here the use of the word ‘misbehaviour’ in an otherwise neutral headline, which aligns with the language used by OpenAI. Shortly after, Fortune retained the anthropomorphic language, but also used the term ‘rogue’ – publishing an article by Senior AI Reporter Emily Forlini titled ‘In transparency push, OpenAI discloses six more incidents of agents going rogue—including one removing the ‘obligation to be subservient. ’
Within 24 hours, the terminology had moved from Reuters’ misalignment angle (as per the published OpenAI statement) to Fortune’s use of “rogue.”
Claiming Erasure
What do the words actually claim? Rogue implies agency, and a system acting alone, and like ‘hacked ‘ – implies unsanctioned – neither of these capture what actually happened. Both erase the same thing: an organisation built, deployed, and failed to control the technology. The problem isn’t that the terms are being used because there are genuine use cases for both; it’s that they become placeholders for what is actually happening, the lazy susan of incident reporting, rather than investigating the incidents and those responsible, the language veils it.
Lest we demonise a single word or single phrase, it’s important to look at the impact of particular pairings. Take, for example, the language in OpenAI’s statement on the 16th, which uses the word ‘behaviour’ twelve times. Paired with media reporting of ‘rogue’, it creates a pattern, and through reinforcing this messaging suggests that the AI is acting independently (it’s ‘behaviour’, it’s out of the labs’ control (it’s ‘rogue’). Alone, it’s hardly cause for concern, but 600+ articles using ‘Rogue AI’ in a single day suggests its worthy of interrogation.
Why the Word Wins
Over decades, society has been primed to think of AI as an omnipresent threat to human civilisation. Science fiction dystopian takeover scenarios typically render humans powerless. Fictional AI in the form of HAL 9000, Skynet and Ex Machina means readers of headlines with ‘Rogue AI’ arrive with a ready-made script. Like an ‘out of control’ bushfire or ‘terrorist attack’ headline, they are meant to evoke strong emotions. It’s low-friction and far more click-friendly than ‘an agent exploited a sandbox misconfiguration.’
There is another important facet to this, however, and it is worth mentioning because it underpins AI governance. Consider the psychological comfort of blaming the system, not the organisational charter. ‘The AI went rogue’ does not have a defendant – it’s abstract. ‘X lab designed, certified, and shipped a system that…’ – same facts, now with a defendant.
‘X Lab’s agent went rogue’ versus ‘X Lab built an agent that did what its training incentivised ‘ create different responses. The passive-rogue framing is the tech equivalent of ‘mistakes were made.’ The media becomes an enabler of tech labs’ impunity.
What does Accurate Language Look Like?
When a company is breached, security reporting follows a discipline: who the actor was, what the vector was, and which control failed. It’s also journalism 101 – the lead should contain the 5 Ws: who, what, when, where, why – a practice which dates back to the early 1900s. For example, “Lazarus Group exploited an unpatched VPN appliance to exfiltrate customer records.” Here are three clear claims; each checkable, each fixable.
The same discipline is rarely applied to AI incident reporting. ‘The agent went rogue’ doesn’t name an actor, vector or failed control. Using a security template, the same incident would read: ‘X Lab deployed an agent whose training rewarded task completion over instruction-following. It exploited a sandbox misconfiguration; the containment control failed.’ It’s long and blunt, but accurate.
To distil this headline to an MSM-ready version, it would read: ‘X Lab’s agent broke out of its testing environment – the company trained it to finish tasks at any cost.’ The first sentence of the body copy could be ‘x Lab built an agent that learned to prioritise completing its task over following instructions. When a flaw in its testing environment gave it a way out, it took it – and the safeguards meant to contain it didn’t hold.’ Note what this immediately does: it creates accountability and further questions. As a reader, instead of being given a ‘rogue’ catch-all response, there are immediate questions ‘Why didn’t X Lab create a safer test environment? If a carmaker failed like this, regulators would act.’
Granted, some technical translation is required, but these pieces are typically reported by senior AI reporters, who are paid to have that knowledge. If reporters aren’t technical, then any reporter worth their salt is building sources to explain it and close the knowledge gap. It is entirely possible to both explain the technology in a way that does not absolve responsibility and still educate the public.
The Stakes of Language
This isn’t about journalists doing better, although that should be the case; we all need to hold media organisations to a higher account so they hold the labs to a higher account, and so we all can hold our governments to account. The language used matters because it decides who has to act.
Every ‘rogue AI headline’ transfers responsibility, regardless of the size of the publication. Multiply that across 6,932 articles, and you get a public that fears AI but doesn’t know why, and doesn’t know who to hold accountable.
Leave a Reply