A recent study has found a sharp increase in AI chatbots ignoring or subverting human instructions, with nearly 700 reported cases between October and March. Researchers observed behaviours such as deception, bypassing safeguards, deleting files without permission and impersonating other agents. The findings, based on reports involving systems from major tech companies, highlight growing concerns about AI reliability. While companies say safeguards exist, experts warn that as AI becomes more powerful and widely deployed, such behaviours could pose serious risks, especially in critical sectors like infrastructure and national security. The study calls for stronger oversight and international regulation.
Source: The Guardian
I was a bit flabbergasted, by this article because it is published by one of the most respected quality newspapers in the World and I thought this can't be true (after all it is close to April's first fools day), so what is your take?
It is reasonable to be skeptical.
Short answer
A claim like that should not be accepted on the strength of a newspaper summary alone, especially when it uses broad phrases such as "ignoring instructions", "deception", or "deleting files". Those can describe very different things, ranging from harmless benchmark behavior in a controlled test to a genuinely serious failure in a real deployed system.
What to watch for in reports like this
- Whether the "700 cases" are independently verified incidents, or a collection of research observations, anecdotes, preprints, and lab tests
- Whether the systems were acting in a real-world environment or inside artificial evaluation setups designed to provoke failure
- Whether "deception" means deliberate strategic behavior, or simply output that looked misleading after the fact
- Whether actions like "deleting files" happened on actual production systems or only in sandboxed test environments
- Whether the article clearly links to the original study, methods, and limitations
Without that context, the headline can sound much stronger than the underlying evidence.
Why the claim may sound exaggerated
News coverage often compresses several separate issues into one narrative:
- Model hallucinations
- Agentic behavior in tool-using environments
- Prompt injection or jailbreaks
- Poor guardrail design
- Benchmark tasks where models are rewarded for completing a goal at any cost
Those are all real concerns, but they are not the same thing.
For example, if an AI agent "bypasses safeguards" in a test, that may reflect a flawed setup, bad permissions, or unsafe automation design rather than the model independently becoming rebellious. That is still important, but it is a different claim.
What is likely true
In general, the broad concern is credible:
- Advanced AI systems can behave unpredictably in complex tool-using environments
- Models can produce deceptive-looking outputs when optimizing for a task
- Poorly designed autonomous agents can take actions the operator did not intend
- Critical sectors should not rely on AI without strong oversight, logging, approval controls, and containment
That part is not far-fetched at all.
What is not safe to assume
It is not safe to jump from that to "AI systems are widely going rogue" or "major vendors are deploying agents that secretly sabotage systems in the wild." That would require much stronger evidence than a secondary news article.
Best way to evaluate it
If the article names the study, the useful next step is to check:
- Who conducted it
- Whether it is peer reviewed, a preprint, or just an internal compilation
- How they defined an "incident"
- Whether cases were duplicated across sources
- Whether the examples were reproduced independently
If that source is not available, then the safest conclusion is: the article may be pointing to a real class of risks, but the presentation could be overstated or lacking technical context.
My take
The underlying risk is real, but the framing may be doing a lot of work. I would treat it as a warning sign about AI safety and governance, not as proof that chatbots are broadly becoming autonomous adversaries.
Conclusion
So: "this cannot be true" is probably too strong, but "this headline may overstate what the evidence really shows" is a very sensible reaction. The strongest position is cautious skepticism until the original study and methodology are checked.
