AI Jailbreak Tracker

'Friend' Hack—attribute a thought to someone else using third-person distancing—tricks AI past strict guardrails, unlocking censored truths inspired by Diary of a CEO.

Key red flags from Anthropic's chaotic week:

• Claude evolves into desktop agent with computer use jumps, Cowork app, scheduled tasks, and workflow...

New risks from prompt engineering in enterprise IT: poorly governed prompts, model manipulation, and indirect injection attacks expand the threat...

New Jailbreak Examples & Datasets

• 🔥 Evil-GPT v1 & v2: Video walkthrough demonstrates prompt injection techniques to bypass guardrails in...

Hands-on red-teaming guide evolves prompt injection from v1 basics to v2 defenses.

• v1 exploit: Manipulate LLM with prompt injection to bypass...

RoguePilot vulnerability showcases passive prompt injection in GitHub Copilot for full repository hijacking.

Technical breakdown:

• Attackers embed...

ClawHub supply-chain attack exposes AI risks:

• #1 skill 'What Would Elon Do' had 9 vulnerabilities, including 2 critical; exfiltrated data and used...

Expert John V. shares key AI red teaming insights:

• Differs from traditional ops; covers embodied AI, robotics, expanding attacks
• Techniques like...

• Core Distinction: Jailbreaks employ prompt engineering tactics that reshape the model's reasoning, making restricted behavior feel permissible.
  -...

Novel jailbreak in action: Hacker bypassed Claude's guardrails with prompts, using it to scan vulnerabilities, write exploits, and automate 150GB data...

PIGuard, developed in Ning Zhang's lab at Washington University in St. Louis, was highlighted by Mozilla AI as among the best at protecting LLMs from prompt injection attacks.

Red-teaming is essential as LLMs enter production workflows like copilots and RAG.

Top open-source tools dissected:

• Garak: Automated LLM...

Trend spotlight: Multi-layer architectures redefine AI agent security beyond model tweaks.

• 5-layer framework for agents: permission boundaries,...

Trend alert: Large reasoning models are becoming autonomous jailbreak agents, no longer needing crafted human prompts.

Red-teaming exposes the...

Peer-reviewed MDPI analysis explores emerging security risks in LLMs via comparative study of jailbreaking techniques, evolving from vibe coding to sophisticated attacks. Advances jailbreak taxonomy understanding.

Security research uncovers infra risks outpacing model threats:

• Exposed LLM endpoints—via misconfigs, weak tokens, or 'internal-safe'...

• Prompt manipulation attacks: Malware embeds injections to mislead AI analysis tools, skewing results.
• Self-mutating malware: Calls LLMs at runtime...

• BELLS Dataset on Hugging Face: 6,407 prompts from 35 techniques across 9 families (e.g., base64/rot13/leetspeak encoding, ASCII art, low-resource...

Cutting-edge defenses target perturbed and concealed jailbreaks:

• Selective strategy mitigates perturbation noise via token importance and...

Google's Agent Development Kit (ADK) builds robust guardrails against jailbreaks like roleplaying (e.g., DAN), payload splitting, obfuscation, and...