Vol. 16: Can We Really Constrain AI?

Most AI “guardrails” feel like the software equivalent of a baby gate: designed to keep trouble contained, but easily toppled with a determined push. The past year of jailbreaks and “DAN” prompts proved that point all too well.

Two new arXiv papers try to move the discussion beyond duct tape and disclaimers:

• “JADES” introduces a systematic way to score and compare jailbreak risks. Instead of playing whack-a-mole with prompt exploits, it decomposes models into measurable vulnerabilities. Think of it as a stress test for LLM resilience.

• “Governable AI” pushes further—arguing for provable safety even under extreme adversarial conditions. It’s less about slapping on filters, more about building the kind of formal guarantees banks already demand in risk management.

AI is being wired into real workflows: credit decisions, compliance reviews, even board reporting. If models can be tricked into ignoring policies with a clever prompt, regulators won’t care whether it was “alignment drift” or “just a jailbreak.” They’ll care that your controls failed.

The parallel to financial services is clear: a firm doesn’t certify a trading system just because it “works most of the time.” It must be stress tested. It is critical to model worst-case scenarios and demand formal proofs where possible. AI deserves the same discipline if it’s going to move beyond shiny demos and edge cases into core infrastructure.

The question is whether industry will adopt stronger methods now—or wait until enough high-profile AI failures force the issue.

Read the JADES paper here → https://lnkd.in/ekUreisV

hashtag #AlgorithmAndBlues hashtag #AI hashtag #EnterpriseAI hashtag #AIGovernance hashtag #AISafety hashtag #Automation hashtag #RiskManagement hashtag #FinancialServices hashtag #MachineLearning hashtag #ResponsibleAI