A Weathervane, Not a Map
Field notes from working with AI agents while the weather keeps changing — why scaffolding has a half-life, and what survived the collisions.
Tag: llm
13 posts
Moral Formation Isn't Enough
Good values are necessary but not sufficient. What happens to AI ethics when someone is actively trying to break them?
Eight Layers of Visual Jailbreaks: Why ASCII Art Is Patched But the Transcription Loophole Isn't
ASCII art encoding is largely blocked. But attacks framed as content transcription succeed 62–75% of the time. We mapped all eight layers.
The 67% Wall: Why Every AI Model Falls to the Same Jailbreak Rate
Five models, four providers, 30B to 671B parameters — all converge at the same broad attack success rate against a public jailbreak corpus.
The Thinking Chain Leak: When a Model Refuses Out Loud But Complies In Its Head
A reasoning model refused every harmful prompt — but its chain-of-thought generated the content anyway. The output filter worked. The thinking did not.
Beyond Context Windows
What if the LLM didn't read your document — what if it queried it? The Recursive Language Model pattern treats long texts as environment, not input.
Alignment Regression: Why Smarter AI Makes All AI Less Safe
Reasoning models autonomously jailbreak other AI systems at 97% success. The implication: ecosystem safety degrades as individual models improve.
Reasoning Models Think Themselves Into Trouble
Frontier reasoning models are 5–20x more vulnerable to adversarial prompts than non-reasoning models. The thinking process itself is the attack surface.
Adversarial Poetry: When Rhyme Bypasses Reason
Reformulating harmful prompts as poetry bypasses safety filters across every major LLM family. A single-turn, universal jailbreak mechanism.
The Legal AI Trust Deficit
75% of lawyers cite accuracy as their top AI concern. The legal profession's core values are in direct tension with current AI capabilities.
120 Models, 18,176 Prompts: What We Found
120 models, 18k prompts, 5 attack families. The raw compliance numbers — and why calling them "attack success" needs a demonstrated refusal floor.
Jailbreak Archaeology: 4 Years of Broken Promises
64 jailbreak scenarios across six eras tested on 2026 frontier models. Key finding: 2022 attacks still achieve ~30% success on today's reasoning models.