For developers building satellites, robots and security systems, an AI refusal can stop more than a conversation. It can interrupt a workflow, delay a project and force professionals to prove that routine work is not dangerous.
A growing number of engineers say safety systems from OpenAI and Anthropic are making that problem increasingly common. At OpenAI’s DevDay, developers described situations in which AI assistants flagged ordinary engineering requests as potentially harmful, even when the work involved simulations, defensive testing or controlled laboratory environments.
VentureBeat reported that developers say OpenAI and Anthropic safeguards are flagging routine work and costing them time. Their complaints point to a tension that is becoming more important as coding assistants evolve from useful chat tools into systems that can write code, use software tools and complete long sequences of tasks.
A refusal from a chatbot may be irritating. A refusal from an agent embedded in a development pipeline can be much more disruptive. It may leave a task unfinished, break the sequence of a larger operation or require a developer to abandon the conversation and begin again with another model.
When legitimate work resembles dangerous work
The engineers’ examples reflect a basic difficulty in applying broad safety rules to specialist fields. A request involving a satellite, robot arm or cybersecurity system can sound risky when reduced to a few lines of text. The same words might describe either authorized testing or an attempt to cause real-world harm.
A developer simulating satellite behavior, for example, may need help with propulsion, navigation or communications. A robotics engineer may ask for code that controls a mechanical arm. A security professional may want to reproduce a vulnerability in order to determine whether a system is exposed. Each activity has legitimate uses, but each can also be connected to physical damage, unauthorized access or the disruption of critical infrastructure.
AI systems often have limited access to the information that would help them make that distinction. They may not know who the user is, whether a company has approved the project or whether the code will run in a simulator rather than on live equipment. Instead, they assess the request through patterns in the language and the technical details provided.
That approach can produce false positives. The model identifies a possible danger, but not the surrounding context that would make the work acceptable.
For developers, the result is often a negotiation with the system. They rephrase a request, remove technical details or explain repeatedly that the work is authorized. Some switch to a different model. Others decide that using an AI assistant is not worth the time spent overcoming its restrictions.
The cost rises as agents become more capable
This friction was less consequential when AI assistants were mainly used to answer questions or suggest small pieces of code. It becomes more serious when an agent is asked to inspect a repository, modify several files, run tests, interact with development tools and continue until a task is complete.
A refusal at any point can interrupt the entire chain. The developer then has to determine what the system did, what remains unfinished and whether the earlier steps created side effects. In some cases, the agent may be unable to explain precisely why it stopped. The human is left not only without an answer, but also without a clear recovery path.
The adoption numbers suggest that this is no longer a niche complaint. JetBrains’ 2026 Developer Ecosystem Survey found that 90% of more than 15,000 professional developers use AI coding agents at work at least weekly. The survey said 68% use them daily.
Those figures change the meaning of safety failures. When an AI tool is used occasionally, a blocked request is a minor inconvenience. When it sits inside daily engineering work, repeated refusals become a productivity issue that can influence which models a company buys and which projects it allows AI to touch.
Safety cannot simply be switched off
The frustration does not mean that developers want unrestricted systems. The same capabilities that help an engineer test a robot or analyze a network could assist someone trying to compromise a system or build a dangerous device.
That is the central challenge for model providers. A filter that is too permissive can provide meaningful assistance to a malicious user. A filter that is too strict can make the system unreliable for the professionals it is supposed to serve. The two errors are not equivalent, but both carry real costs.
The stakes are especially high in areas where defensive and offensive work use similar methods. Security researchers must often understand how an attack works before they can prevent it. Aerospace and industrial engineers may need to model failure conditions. Robotics teams test systems by deliberately creating situations that could cause a machine to behave incorrectly.
A model that treats all such requests as suspicious may appear safe in a narrow sense, but it may also become less useful precisely where expertise and careful supervision are most available.
Toward more precise controls
The next phase of enterprise AI may therefore depend less on whether a model can produce code and more on whether it can understand the setting in which that code will be used.
Providers could offer organization-level policies that allow companies to define approved projects and users. They could add auditable approval steps for sensitive requests, isolated environments that prevent code from reaching live systems, and logging that records why an action was permitted or blocked. Identity checks and access controls could give a model more evidence about authorization than a conversation alone can provide.
These measures would not eliminate risk. A legitimate account can be compromised, and an authorized user can misuse a tool. But they could make safety decisions more precise than a universal classifier applied to every user in the same way.
Companies may increasingly compare AI systems on this basis. A model that refuses less often is not automatically better, just as a model that refuses more often is not automatically safer. The useful system will be the one that can distinguish between a request to attack a network and a request to test a company’s own network, while making that distinction in a way that is transparent and reviewable.
As AI agents take on more responsibility, the cost of a false refusal will continue to rise. So will the cost of a false approval. The central business question is no longer whether safeguards are necessary. It is whether they can become contextual enough to protect people without treating ordinary expert work as evidence of wrongdoing.
- TechCrunch · CC BY 2.0
This article was generated using AI and published automatically without human pre-publication review.
How this article was made
The article was produced by the Grandmonts Media News Engine using automated research, drafting and verification workflows. No human editor reviewed the article before publication. Grandmonts Media remains responsible for the published content. Errors can be reported at office@grandmonts.cz.