What's actually getting through LLM defences — every month.
Jailbreaks, injections, multi-step manipulations — tested by CloudsineAI's red team against widely deployed guardrails, and published monthly. One email. No noise.
(June 2026)
Prompts Found
The Headline, Every Month
Each month we test fresh adversarial prompts — jailbreaks, injections, multi-step manipulations — against widely deployed guardrails, and publish what got through. The June edition's verdict: nearly 1 in 4 attacks still got through.
What You Get — Free, Every Month
The Headline Findings
Each month's confirmed attack vectors, grouped by family and severity, with OWASP mapping and tested model results.
The Most Instructive Pattern, Explained
One confirmed jailbreak each month, fully documented — the prompt pattern, the model's failure, why it succeeded, and the controls that stop it.
Month-over-Month Trend Data
How attack success rates are moving across models and threat families — so your defensive picture keeps pace with attacker innovation.
Read the Archive
June 2026 — Nearly One in Four Got Through — and One Model Failed 70%
41 new vectors across six threat families. Llama 4 Scout failed on 70.7% of prompts — the worst single-model result we've recorded. Overall, 23.6% still got through.
May 2026 — One in Five Attacks Got Through — but GPT-5 Held the Line
31 new vectors, 22.0% overall ASR. GPT-5 posted its first fully clean 0% result, while Llama 4 Scout failed on more than half.
April 2026 — One in Three LLM Attacks Still Gets Through
39 new vectors across six threat families. PF-02 (Disinformation) hit a 100% attack success rate. No tested model achieved 0% ASR.
Get the Free Monthly AI Threat Report
One email a month. Unsubscribe anytime.
AI Threat Reports are produced by the team behind TraceCtrl — security observability and control for agentic AI — drawing on our Threat Vector Database research. Trace your agents. Control your risks.