SALUCA: Shield Around Local Unprotected Computerized Agents

Artificial intelligence agents are no longer chatbots. They execute code, retrieve documents, manage schedules, callexternal APIs, and act on real-world systems with increasing autonomy. This shift from conversational to autonomousoperation exposes a fundamental gap: the infrastructure that makes agents capable has outrun the infrastructure thatmakes them safe. An AI agent today operates with the power of a networked software system and the trust model of ademo.This paper describes a complete infrastructure-level security stack for AI agents, comprising four technical pillarsdeveloped across a portfolio of seven patent applications. The four pillars are: (1) a cryptographic API forcefield thatdetects prompt injection attacks without reading a byte of conversation content; (2) a persistent memory architecturethat survives context resets while maintaining hash-verified integrity at every read and write; (3) a personalized recallengine that surfaces the right memories at the right moment using topic-keyed routing and adaptive weighting; and (4)a verifiable identity and provenance framework that makes every memory retrievable with a cryptographicallyauditable chain of custody.Three of these pillars are unified under the Tiresias Platform brand: tiresias.watch (injection detection) andtiresias.report (hybrid PII detection with compliance audit) ship as named products, while a third pillar — federatedanonymized threat intelligence — is covered by a separate patent application (the Tiresias Network patent) and isdescribed in this paper as an architectural layer rather than a packaged product. The fourth pillar, memoryinfrastructure, is implemented in the Soul/Elysium memory engine. A fifth invention, the cross-layer orchestrationpatent, acts as the capstone — binding all signals into a single composite risk score, a single configurable enforcementpolicy, and a single hash-chained audit trail that satisfies GDPR, HIPAA, and SOC 2 requirements without requiringseparate data collection.The unifying thesis is this: AI agents need not be secured by reading and filtering their content. They can be secured atthe infrastructure level — through cryptographic guarantees, structural integrity verification, and adaptive riskcorrelation — in ways that are blind to content but aware of everything that matters about how that content flowsthrough the system. This paper de AI disclosure. This work was produced with an agentic AI research apparatus operated by Saluca Labs. The apparatus drafted, searched and analysed under direction. Cristian Ruvalcaba is the human author and is accountable for the content. No AI system is listed as an author or contributor, because authorship entails accountability that a model cannot hold; this disclosure is the credit, and it is deliberately the whole of it.