OpenAI and Anthropic Guardrails Are Getting in the Way of Cybersecurity Researchers

On July 23, 2026, TechCrunch published a piece reporting that the guardrails in OpenAI and Anthropic language models are getting in the way of offensive cybersecurity researchers, the specialists who hunt for still-unknown vulnerabilities and develop tools to exploit them.

Why AI filters get in pentesters' way

OpenAI (ChatGPT) and Anthropic (Claude) models increasingly refuse to help with offensive security tasks, even when a white-hat professional is at the keyboard. According to TechCrunch, the publication surveyed several researchers who professionally hunt for zero-day vulnerabilities and write exploits, and all of them say the same thing: the filters treat a legitimate work task as an attempted hack.

The problem is that a request like "write code that exploits this vulnerability" looks identical to the model no matter who sends it: a company's red team or an attacker. The model sees only the text of the request, not the intent and not the pentest contract.

• Publication: TechCrunch, published July 23, 2026
• Models in focus: OpenAI (ChatGPT) and Anthropic (Claude)
• Who was surveyed: offensive security researchers
• The core complaint: guardrails refuse to help with finding vulnerabilities and writing exploits

Guardrails are safety rules built into a model that block potentially harmful output: instructions for building weapons, malicious code, or ways to bypass security protections. OpenAI and Anthropic position them as a key element of responsible AI, but it is precisely at the border with offensive security that the rules trigger too broadly.

What is the dual-use dilemma?

Offensive cybersecurity is a classic dual-use technology: the same skills and tools that protect infrastructure can also break it. A red team imitates an attacker to find holes before a real adversary does, and for that it needs exactly the same exploits, malware emulation, and defense-evasion techniques.

The guardrails at OpenAI and Anthropic are tuned to block exactly this class of requests, because the labs cannot tell a researcher from a criminal on the fly. As a result, as TechCrunch describes it, security specialists hit refusals where a year or two ago the model readily helped, and they are forced to work around the restrictions or turn to less regulated tools.

The guardrails of

OpenAI and Anthropic are getting in the way of offensive cybersecurity researchers: that is how TechCrunch's July 23, 2026 article frames the core of the problem.

What this means for defense

Overly tight filters hurt more than the offensive side; they slow down defense. If a legitimate red team cannot use a top-tier model for routine work, it loses speed, and the model loses usefulness for an entire professional segment. According to TechCrunch, some of those surveyed are switching to local open-weight models or narrowly specialized security tools that have no corporate filters, at the cost of quality and convenience.

That creates a skew: attackers already work with unrestricted models (open-weight, "uncensored" builds, underground services), while the people with their hands tied are the ones acting legally and under contract. By the logic of TechCrunch's article, excessive caution at OpenAI and Anthropic risks punishing conscientious researchers without stopping unscrupulous ones.

What this means

The debate over guardrails in offensive security is a debate over where the line runs between "safe by default" and "useless for professionals." Until OpenAI and Anthropic learn to tell a legitimate pentest from attack preparation, for example through organization verification or special modes for security teams, researchers will choose the tools that do not block them.