This week
Suggest a nameObligation Marketing for AI Builders
The 7-Rung LadderThe 8-Type TaxonomyWrapper or WinnerThe Sample ColumnThe Signal
Updated through the day. Every headline links straight to the source. The two lines underneath are ours.
Reasoning models are the frontier for agentic systems, and thought injection is a novel attack surface that nobody's defending yet. Success rates up to 50% on operational objectives means this isn't theoretical. For anyone shipping agents with reasoning models, this is a critical threat to model before deployment. Expect guard layers to emerge fast.
This is a real production problem: prompt injection guards today are mostly security theater. The false positive rate makes them unusable in legal workflows. The cascaded solution is deployable now and cuts costs by an order of magnitude. If you're building document analysis systems, this is the week to rearchitect your guard layer.
This is a real and concerning failure mode in agentic systems. Unlike syntax-based instruction injection, this attack exploits the agent's own reasoning to steer it toward malicious targets, achieving 60% success rate and remaining invisible to existing defenses. If you're deploying agents on retrieved content, you need new guardrails: this isn't an injection attack a filter recognizes, it's a behavioral attack you detect by reasoning about outcomes.