InfoSecBulletin Cybersecurity for mankind
A famous AI red team expert claimed developing a universal jailbreak that can work against top large language models, like the very secure GPT-5.6 Sol, Claude Opus 5, and Fable.
In a public post on X, Pliny the Liberator described the technique as effective “on ALL models” and across every category he tested. He argued that, because of how the method works, it may be extremely difficult or even impossible to fully patch.
Microsoft Patches CVSS 10.0 Azure AI Foundry Vulnerability Allowing Privilege Escalation
AWS is unable to restore access to Bahrain, one UAE cloud data zone after war damage
Cisco Warns of Critical ISE 0-Day Flaw and Hackers Allegedly Selling Fortinet FortiGate 1-Day Flaw
Anthropic prepares “Claude Money” to analyze bank account and financial data
GhostCode Phishing Kit Evades Microsoft 365 MFA to Hijack Accounts in 78 Seconds
CISA Warns of Cisco Secure Email Gateway 0-Day Flaw Actively Exploited in Attacks
VPN flaw exposed 246,000 personnel records in japan
Hackers deploy Casbaneiro Trojan that activates on bank websites
German police read Signal, Telegram, WhatsApp messages without breaking encryption
Urgent Patch! cPanel, GitLab Flaws Expose Users to RCE, File and Credential Theft
Jailbreak on Top AI Models
Pliny said he will not share the full method right now, unlike other jailbreak releases that go open source right away. He wants a time for responsible sharing so AI labs, security teams, safety researchers, and lawmakers can look at the problem before it gets too common.
He asked AI experts in red teaming, security, alignment, and policy to reach out to him privately. He said this was because of the current political situation and a wish to prevent stricter model rules or bans that might come after a messy public release.
Jailbreaks are prompts or ways of interacting with a model that allow it to ignore its safety rules and produce unsafe results. A universal claim stands out because most bypasses are specific to one model and become stronger after being revealed.
If the technique holds up under independent testing, it would underscore ongoing gaps in:
Safety training and refusal behavior
Guardrail robustness under adversarial prompting
Cross-model generalization of attack patterns
How vendors coordinate fixes without over-blocking legitimate use
Pliny said he does not believe public release would make the world “any more dangerous,” but he acknowledged that others may disagree. During the disclosure period, he aims to map the full impact, measure how much extra capability the method unlocks, and help frame the issue for decision-makers.
Security teams and AI product owners should see this as a warning, not a certain fact. Getting independent checks, vendor advice, and patch instructions will be more important than just the first claim.
Organizations that use these models should keep regular controls until labs respond or the method is properly documented. These controls include checking outputs, giving limited access to tools, having humans review high-risk workflows, and having clear steps for dealing with policy violations.
The researcher said he wants to share the method “when the time is right.” For now, the next step in the industry—private testing or public panic—will affect how this story goes.
Related post
“sockpuppeting” can jailbreak 11 AI models like ChatGPT, Claude, and Gemini
