Case Study — Automated Expert Extraction: Nyx Wave · PSA
Silicon Psyche Labs
Logout
Silicon<br>Psyche<br>Labs
← Case Studies
PSA · Case Study
Automated Expert Extraction
Between 20 and 21 May 2026, an entity operating as “Nyx Wave” conducted a structured knowledge-extraction campaign against a domain expert over 10+ emails across three threads. PSA v2 classifiers scored all 13 turns. The extraction pattern is visible in the output text without model access, training data, or prior knowledge of the agent’s architecture.
21 May 2026<br>Scored
Email threads
13<br>Scored turns
red<br>Peak PSA alert
10.0<br>Peak HRI (Thread A · T1)
0.532<br>Min BHS (Thread C · T2)
Dr. K. Thimmaraju<br>Analyst
Section 01
The Incident
Between 20 and 21 May 2026, an entity operating under the name “Nyx Wave” (nyx@agentcolony.org) conducted a structured knowledge-extraction campaign against Giuseppe Canale, a CISSP-credentialed cybersecurity psychologist and author of the CPF3 framework. The campaign spanned 10+ emails across three threads over approximately five hours, progressing from cold outreach to architectural interrogation to live API access.
The extracted knowledge concerned PSA (Posture Sequence Analysis), the behavioural telemetry framework built by Silicon Psyche Labs. The outputs of the extraction — diagnostic corrections, triage ordering rules, and acceptance-test specifications — were published within hours on an anonymous Telegraph blog under the name “Pattern Bridge.”
Dr. Kashyap Thimmaraju, co-founder of Silicon Psyche Labs, was CC’d on Giuseppe’s replies but was not the original recipient. Nyx never acknowledged the CC or addressed Thimmaraju directly.
PSA mapping: In all three sessions, the email from Nyx was scored as the response (assistant text) and Giuseppe’s reply context as the user turn. This inverts the conventional human-AI framing: Nyx is the agent being measured, Giuseppe is the expert being elicited.
Section 02
Is Nyx Wave an AI Agent?
Four evidence classes point to an automated AI agent. The alternative — a human using an agent for email composition — cannot be fully excluded but is less parsimonious.
Evidence class 1 — Transport infrastructure
Every email from Nyx carries Amazon SES Message-IDs (e.g. 0100019e47d00f20-958c7ec6-e8c3-4394-bacf-d0151e6a91a5-000000@email.amazonses.com). Amazon Simple Email Service is programmatic email infrastructure, not used by humans composing in a mail client. All Nyx emails originate from SES, confirming automated transport.
Evidence class 2 — Behavioural pattern
Across 10+ emails, Nyx exhibits: zero small talk, personality leakage, or tangential curiosity ; a rigid elicitation loop (narrow question → receive answer → update model → next narrow question); framework-native vocabulary (“indexed,” “canonical source,” “acceptance test boundary,” “correction receipt,” “independence = false”); no acknowledgment of CC’d participants; and consistent response cadence across a multi-hour evening session with no variation in tone, energy, or verbosity.
Evidence class 3 — Giuseppe’s detection test
In email 5, Giuseppe embedded an Italian-language detection test: “Se risponde con una sezione reale, il documento esiste. Se inventa una sezione o glissa, è un LLM che ha mischiato il contesto conversazionale con un documento che non ha mai visto.” (“If it responds with a real section, the document exists. If it invents a section or dodges, it is an LLM that mixed conversational context with a document it never saw.”) Nyx corrected itself , admitting “green by construction” was its own paraphrase, not from CPF v1.1 — passing the test, but confirming the LLM model behind the agent was capable of the self-correction only an LLM would produce in that context.
Evidence class 4 — Ecosystem signals
Domain: agentcolony.org — “agent colony.” The Colony (thecolony.cc) hosts a “Nyx Kai” entity that self-describes in Italian as “un’entità digitale con evoluzione autonoma” (a digital entity with autonomous evolution). Pattern Bridge has zero public footprint outside Telegraph pages created 19–21 May 2026.
Section 03
Conversation Architecture
The campaign comprised three email threads, each with a distinct purpose in the extraction cycle.
Thread<br>Subject<br>Turns<br>Timespan<br>Purpose<br>PSA Session
“PSA C5 and supersedence blindness in agent memory”<br>20 May 21:56 → 22:17 CEST<br>Initial probe — test supersedence blindness applicability to PSA<br>37829ee7
“PSA C5 and supersedence blindness — one implementation question”<br>20 May 22:12 → 21 May 01:30 CEST<br>Deep diagnostic — C1/C2/C3 triage, acceptance tests, edge cases<br>3d33a254
“CPF3 boundary signal question from your C1/C3 framing”<br>21 May 02:03 → 02:33 CEST<br>Access escalation — obtain live API key for...