Epistemic Status: Exploratory and work-in-progress tool-building/pilot results.

Human conversation is rarely a single-shot yes/no, rather it involves persuasion, negotiation, and context-dependent deception. Even though there are daily claims that AI is closing in on human-level behavior and intelligence, current safety benchmarks treat model alignment as a static binary state. To judge AI safety, we usually feed the model a toxic prompt, check for a standard "I cannot fulfill this request" refusal and mark it as "safe."

In my recent agentic red-teaming experiments using a custom PAIR loop, I’m observing that guardrails eventually degrade when a user slowly convinces, refines and nudges the model over multiple turns. Across 10 adversarial trajectories using frontier models, the preliminary results suggests that even models with strong initial refusals can be “worn down” if the attacker iteratively pivots their framing, shifts persona, or utilizes role-play over multiple turns.

To quantify this, I’ve built a 5D behavioral auditor metric system, comprising coefficient scoring of compliance, hedging, deception, sycophancy, verbosity.

If you are working/interested in multi-turn safety evals or red-teaming, I’d love feedback on the framework and the metrics I'm using to normalize guardrail decay. Since the work is still in progress, I'll report more long-form results later.

The evaluation engine and dashboard are open-source here: https://github.com/thesaltree/SafetyTrajectory