Offensive cybersecurity researchers report that safety guardrails from OpenAI and Anthropic are restricting their ability to find vulnerabilities and develop security tools. The limitations are creating friction for legitimate security work.
Cybersecurity researchers who hunt for unknown vulnerabilities face growing obstacles from AI safety measures designed to prevent misuse. OpenAI's and Anthropic's guardrails restrict access to code generation, exploit development, and other technical capabilities these professionals depend on.
The issue centers on distinguishing defensive research—identifying weaknesses to fix them—from malicious activity. Researchers say guardrails often block legitimate security work without clear appeal processes or exceptions for credentialed professionals.
This creates a broader tension: AI companies prioritize safety by default, but overly broad restrictions can hamper the security community's ability to protect systems. Researchers lack consistent ways to demonstrate their credentials or request temporary access for authorized work.
Both companies have implemented some researcher programs, but coverage remains limited. The cybersecurity community is calling for more nuanced policies that account for legitimate offensive security research without compromising safety.
A race condition vulnerability in the Linux kernel's XFS filesystem allows local attackers to overwrite protected files and gain root privileges. The flaw, tracked as CVE-2026-64600, has remained unpatched for nine years.
Intensified enforcement against online scam operations in Cambodia is displacing criminal networks to Sri Lanka, where authorities have arrested over 1,000 people in 2026.
European drug traffickers are leveraging AI-boosted chemical synthesis to create designer drug precursors that circumvent existing product blacklists, according to an EU agency warning.
Domain registrar Namecheap handed over a customer's account to an unverified third party after a password reset request, raising security concerns for a 13-year customer.