Researcher Claims Jailbreak for GPT-5.6, Claude Opus 5, and Fable AI Models

Key Takeaways

• A prominent AI red teamer, Pliny the Liberator, claims to have developed a “universal jailbreak” affecting major large language models.
• The alleged jailbreak impacts top-tier models, including GPT-5.6 Sol, Claude Opus 5, and Fable.
• The researcher is withholding the full technique for a responsible disclosure period, seeking collaboration with industry experts.
• If validated, this could expose significant vulnerabilities in AI safety mechanisms and prompt a re-evaluation of current guardrail strategies.

A leading figure in AI red teaming has announced the development of what he describes as a universal jailbreak, capable of bypassing the safety protocols of several advanced large language models (LLMs). The researcher, known as Pliny the Liberator, asserts that this technique is effective across a wide array of models, including highly protected flagship systems like GPT-5.6 Sol, Claude Opus 5, and Fable.

In a public statement shared on X, Pliny characterized the method as universally applicable, working “on ALL models” and across every category he subjected to testing. He further suggested that the fundamental nature of this technique might render it exceptionally challenging, if not impossible, to fully mitigate through conventional patching methods.

Jailbreak on Top AI Models

Unlike many vulnerability disclosures that are immediately made open source, Pliny has chosen to temporarily withhold the full details of this technique. His stated objective is to facilitate a responsible disclosure window, allowing AI laboratories, red teamers, safety researchers, and policymakers to thoroughly examine the implications before the method becomes widely known.

He extended an invitation to experts in AI red teaming, security, alignment, and policy to engage with him privately. Pliny articulated that this measured approach was influenced by the prevailing political and regulatory landscape, aiming to prevent more stringent model restrictions or outright bans that could arise from an uncontrolled public release, as seen in his post on July 24, 2026.

Jailbreaks typically involve specific prompts or interaction patterns designed to circumvent an LLM’s safety filters, compelling it to generate content that is normally prohibited or carries high risks. The claim of a “universal” jailbreak is particularly noteworthy, as most known bypasses are model-specific and are usually addressed and hardened by vendors following their disclosure.

Implications for AI Safety and Security

Should independent validation confirm the efficacy of this alleged universal technique, it would highlight persistent deficiencies in several critical areas:

• The effectiveness of current safety training and refusal mechanisms in LLMs.
• The resilience of guardrails when confronted with sophisticated adversarial prompting.
• The tendency for attack patterns to generalize across different AI models.
• The complexities vendors face in coordinating fixes without inadvertently restricting legitimate model functionalities.

Pliny expressed his belief that a public release of the method would not inherently make the world “any more dangerous,” though he acknowledged that this perspective might not be universally shared. During the ongoing disclosure period, his aim is to comprehensively map the potential impact, quantify the additional capabilities unlocked by the method, and assist in framing the issue for relevant decision-makers.

For security teams and AI product owners, this announcement should be regarded as an early warning rather than definitive proof. The eventual independent validation, official vendor advisories, and any coordinated patching guidance will hold greater weight than the initial claim itself. Until AI labs issue formal responses or the method is properly documented through established channels, organizations relying on these models should maintain their existing security controls.

The researcher indicated his readiness to share the method “when the time is right.” For now, the industry’s response—whether through private collaborative testing or public concern—will determine the trajectory of this evolving situation.

What You Should Do

• Maintain Output Monitoring: Continue rigorous monitoring of LLM outputs for any signs of policy violations or unusual behavior.
• Implement Least-Privilege Access: Ensure that tools and systems interacting with AI models operate with the absolute minimum necessary permissions.
• Mandate Human Review: For workflows involving high-risk or sensitive data, implement human oversight and review of AI-generated content.
• Establish Clear Escalation Paths: Define and communicate clear procedures for reporting and escalating any suspected policy breaches or security incidents related to AI model usage.

Disclaimer: HackersRadar reports on cybersecurity threats and incidents for informational and awareness purposes only. We do not engage in hacking activities, data exfiltration, or the hosting or distribution of stolen or leaked information. All content is based on publicly available sources.