← All news
BleepingComputer

Anthropic Updates Claude Models, Loosening Restrictions on Security Tasks While Boosting Bug-Finding Skills

Anthropic Updates Claude Models, Loosening Restrictions on Security Tasks While Boosting Bug-Finding Skills

Anthropic has begun rolling out two updated Claude models, with one version going to select partners and another reaching general users. Alongside broader improvements to coding and reasoning capabilities, researchers observed notable shifts in how the models handle cybersecurity-related prompts: the newer versions appear to decline fewer security-focused tasks than prior iterations, while also demonstrating improved ability to locate software vulnerabilities.

This development reflects a broader trend across large language models used in security contexts. AI assistants are increasingly deployed by both defenders and researchers to assist with code review, vulnerability discovery, and penetration testing tasks. Model providers typically build in guardrails that cause an assistant to refuse requests it classifies as potentially malicious, such as writing exploit code or detailing attack techniques. As models improve at technical reasoning, however, providers face a persistent tension: tightening restrictions can block legitimate security work such as bug bounty research or defensive tooling, while loosening them raises the risk that the same capabilities could be misused to accelerate the discovery of exploitable flaws by threat actors.

The dual-use nature of AI-assisted vulnerability research is a growing concern for the security community. Tools that get better at spotting bugs are valuable for defenders patching systems before attackers find the same flaws, but equally valuable to adversaries probing for weaknesses to exploit. As AI coding assistants become embedded in development pipelines and security tooling, organizations should expect continued shifts in what these models will and won't do, often without clear advance notice.

Security teams evaluating or integrating AI assistants into development, code review, or vulnerability research workflows should track model version changes closely, test how updates affect both false refusals and the quality of security-relevant outputs, and avoid assuming a model's guardrails remain static over time. Where AI tools are used to probe internal systems for flaws, findings should still be validated through standard responsible disclosure and remediation processes rather than relied upon as a sole source of truth.

#ai#llm#vulnerability-research#anthropic#ai-security

Source: BleepingComputer

Read the original report ↗

This is an original summary written by CyberSkii for reference. All rights to the underlying reporting remain with the publisher linked above.

Need help? Contact us
Anthropic Updates Claude Models, Loosening Restrictions on Security Tasks While Boosting Bug-Finding Skills — CyberSkii News