LinkedIn publication text
AI Jailbreak: from the iPhone to language models
In 2007, hackers “broke the iPhone out of jail” to install unauthorized software. Fifteen years later, the term took on new life: today, jailbreaking is any technique that gets an AI to produce what it was trained to refuse.
The difference is profound. On a phone, the attack exploits code flaws. In AI, the weapon is language itself. That is why the problem has no definitive fix, and prompt injection is risk No. 1 in the OWASP Top 10 for LLM applications.
I prepared this infographic with a 7-step study:
- Fundamentals: why alignment fails.
- Taxonomy: 8 attack families, from social engineering to GCG.
- Prompt injection and agents: the “lethal trifecta”.
- Defenses in 4 layers: model, filtering, architecture, and operations.
- Evaluation: ASR, unwarranted refusals, and benchmarks.
- Governance: a matrix with OWASP, MITRE ATLAS, NIST AI RMF, and ISO/IEC 42001.
- Continuous operation: monitor, respond, and learn.
Three lessons for those working in GRC:
- LLM security is a probabilistic control, not a deterministic one.
- Design assuming some attack will work, and limit what it can do.
- The frameworks already cover the topic; the challenge is applying them with rigor.
A contrast that says a lot: in 2010, the U.S. exempted smartphone jailbreaking from the DMCA's ban; the EU AI Act, approved in 2024, has required adversarial testing of AI models with systemic risk since August 2025. In 15 years, the practice went from the hacker underground to a regulatory requirement.
It is 19 slides, with a timeline, a workflow, a glossary of 20 terms, and a verified bibliography of 37 references. Swipe to see.
How is your organization handling jailbreak risk in AI agents?
Aridio Silva
Independent Researcher
ORCID: 0009-0008-2411-6995
Creator of SGAEIA
aridiosilva.com/sgaeia
© Copyright Aridio Silva @aridiosilva · October 2026. All rights reserved.