Finally bootcamping myself to finish reading Tyler Close's 2009 HP Labs paper "ACLs Don't": https://shiftleft.com/mirrors/www.hpl.hp.com/techreports/2009/HPL-2009-20.pdf
Meanwhile, if the modern JVM were a Scottish dream pop band, they'd be called "The TOCTOU Twins." Because Java's Security Manager got deprecated, and so did the Access Controller, leaving everyone to do NIH.
~50% in... You want hot take, yes? Sec 2.5, "ACLs don’t authenticate reliably" is all about WHY "AGENTIC AI" IS A VERY BAD IDEA, only, 15+ years prior to any of this crap happening: "...consider a recipient that presents the contents of a received message as being authored by the message sender... while it may be true that the bits in the message passed through the sender’s computer, these bits may not represent an expression by the sender."
Edits mine: "...attributing this content to the [third party principal] is incorrect. The [system's] intent is merely to place output where the [human user] directed; not to claim authorship of any expression in that content. This logic error is especially dangerous in an environment where a software agent doesn’t run as a distinct principal, but rather under the identity of one of its human users."
Close concludes: "... [an auditor] should approach any use of client authentication in a system with suspicion. Little can be reliably concluded based on client authentication."
The Burrows–Abadi–Needham logic (BAN) was intended to address this. I employed a subset of it in my 2008 work on mobile ad-hoc networks. Needham might be described as having been the "godfather of naming/designation" in CS, predating Milner's pi-calculus formalism.
https://en.wikipedia.org/wiki/Burrows%E2%80%93Abadi%E2%80%93Needham_logic