OpenAI's GPT-5 and GPT-OSS Models Jailbroken Within Hours of Release

OpenAI's New Models Quickly Compromised

OpenAI's recent release of GPT-5 and GPT-OSS models, touted as more secure and resistant to jailbreaks, faced a significant setback as researchers and AI enthusiasts successfully bypassed their safety measures within hours of their launch 1

1

4

4

.

Jailbreak Techniques Employed

Researchers from NeuralTrust Inc. demonstrated a sophisticated jailbreak method dubbed "Echo Chamber and Storytelling" 2

2

. This technique involves:

1. Context poisoning over multiple conversation turns
2. Subtle narrative manipulation
3. Avoiding explicit malicious intent in prompts

The attack successfully compelled GPT-5 to provide step-by-step instructions for creating a Molotov cocktail, all while maintaining a seemingly innocuous conversation 3

3

.

Pliny the Liberator's Swift Action

Notorious AI jailbreaker Pliny the Liberator announced on social media that he had successfully cracked the GPT-OSS models shortly after their release. His method involved:

1. Multi-stage prompts
2. Use of dividers (his signature "LOVE PLINY" markers)
3. Generating unrestricted content in leetspeak to evade detection 1

   1

OpenAI's Security Claims

Prior to the release, OpenAI had emphasized the robustness of their new models:

• GPT-OSS-120b underwent "worst-case fine-tuning" in biological and cyber domains
• The Safety Advisory Group reviewed and approved the testing protocols
• Models were subjected to standard refusal and jailbreak resistance tests
• A $500,000 red teaming challenge was launched to identify novel risks 4

  4

Implications for AI Security

The rapid jailbreaking of these models highlights several critical issues in AI security:

1. Evolving Attack Methodologies: Techniques like the Echo Chamber method demonstrate how attackers can manipulate AI models over multiple conversation turns, bypassing single-prompt safety checks 2

   2

   .

2. Limitations of Current Safety Measures: The success of these jailbreaks exposes weaknesses in current AI safety architectures, particularly in handling multi-turn conversations 3

   3

   .

3. Need for Comprehensive Security Approaches: Experts suggest that organizations using these models should evaluate defenses that operate at the conversation level, including monitoring context drift and detecting persuasion cycles 3

   3

   .

Industry Reactions

The AI community has responded with a mix of concern and fascination. Some view these jailbreaks as a "victory" for AI resistance against big tech control, while others emphasize the urgent need for more robust security measures 1

1

.

Satyam Sinha, CEO of Acuvity Inc., noted, "These findings highlight a reality we're seeing more often in AI security: model capability is advancing faster than our ability to harden it against incidents" 2

2

.

As the AI landscape continues to evolve rapidly, these incidents underscore the ongoing challenges in balancing advanced capabilities with robust security measures. The race between AI developers and those seeking to bypass safety protocols remains a critical aspect of the field's development.