BEGIN:VCALENDAR VERSION:2.0 PRODID:IEEE vTools.Events//EN CALSCALE:GREGORIAN BEGIN:VTIMEZONE TZID:America/Los_Angeles BEGIN:DAYLIGHT DTSTART:20270314T030000 TZOFFSETFROM:-0800 TZOFFSETTO:-0700 RRULE:FREQ=YEARLY;BYDAY=2SU;BYMONTH=3 TZNAME:PDT END:DAYLIGHT BEGIN:STANDARD DTSTART:20261101T010000 TZOFFSETFROM:-0700 TZOFFSETTO:-0800 RRULE:FREQ=YEARLY;BYDAY=1SU;BYMONTH=11 TZNAME:PST END:STANDARD END:VTIMEZONE BEGIN:VEVENT DTSTAMP:20261015T190547Z UID:FCDBC323-1B81-41CF-A772-87C3EDDCD32B DTSTART;TZID=America/Los_Angeles:20261105T190000 DTEND;TZID=America/Los_Angeles:20261105T200000 DESCRIPTION:As enterprises operationalize generative AI and autonomous agen tic workflows\, compute expenditure has evolved from an accounting metric into an active security and reliability boundary. Building upon research p resented in Observability for Large Language Models (Apress / Springer Nat ure) and patented methodologies in compute optimization and hallucination detection\, this session provides an architectural deep dive into the dual imperatives of GenAI FinOps and compute protection.\n\nThe talk deconstru cts the emerging threat landscape of LLMJacking and Denial-of-Wallet (DoW) attacks—where compromised credentials and unconstrained agent loops are exploited by threat actors to divert enterprise LLM capacity to unauthori zed workloads. We examine:\n\n- Algorithmic Cost Reduction: Implementing t iered routing frameworks between Small Language Models (SLMs) and frontier models\, semantic caching layers\, and vector search pruning to reduce re dundant token overhead by up to 60%.\n- Zero-Standing Quotas & Cryptograph ic Identity: Applying SPIFFE/SPIRE-based short-lived workload attestation to model inference endpoints\, preventing compromised API keys from allowi ng unbounded cloud spend.\n- Automated Circuit Breakers: Establishing sub- minute anomaly tripwires to isolate rogue recursive agent loops and anomal ous token velocities in production.\n- Return on Autonomy (RoA) Modeling: Translating raw token consumption into boardroom-defensible metrics that b alance MTTR reduction against compute overhead.\n\nKey Attendee Takeaways\ n\n- Threat Mitigation Blueprint: Actionable telemetry patterns to identif y and neutralize LLMJacking attempts before enterprise token quotas are ex hausted.\n- Production FinOps Patterns: Architectural designs for context- window compression and dynamic inference routing that immediately decrease operational cloud spend.\n- Executive Governance Framework: Clear policie s for chargebacks\, environment-scoped spend tripwires\, and engineering-v ersus-token ROI trade-offs.\n\nSpeaker(s): Ankush Sharma\, \n\nRoom: 108\, Bldg: Alameda Hall\, Santa Clara University\, Santa Clara\, California\, United States\, Virtual: https://events.vtools.ieee.org/m/585464 LOCATION:Room: 108\, Bldg: Alameda Hall\, Santa Clara University\, Santa Cl ara\, California\, United States\, Virtual: https://events.vtools.ieee.org /m/585464 ORGANIZER:edcheng@ieee.org SEQUENCE:42 SUMMARY:Autonomous Infrastructure Economics: Algorithmic FinOps\, Compute O ptimization\, and Mitigating LLMJacking URL;VALUE=URI:https://events.vtools.ieee.org/m/585464 X-ALT-DESC:Description: <br /><p>As enterprises operationalize generative A I and autonomous agentic workflows\, compute expenditure has evolved from an accounting metric into an active security and reliability boundary. Bui lding upon research presented in <em>Observability for Large Language Mode ls</em> (Apress / Springer Nature) and patented methodologies in compute o ptimization and hallucination detection\, this session provides an archite ctural deep dive into the dual imperatives of GenAI FinOps and compute pro tection.</p>\n<p>The talk deconstructs the emerging threat landscape of <s trong>LLMJacking</strong> and <strong>Denial-of-Wallet (DoW)</strong> atta cks&mdash\;where compromised credentials and unconstrained agent loops are exploited by threat actors to divert enterprise LLM capacity to unauthori zed workloads. We examine:</p>\n<ul>\n<li><strong>Algorithmic Cost Reducti on:</strong> Implementing tiered routing frameworks between Small Language Models (SLMs) and frontier models\, semantic caching layers\, and vector search pruning to reduce redundant token overhead by up to 60%.</li>\n<li> <strong>Zero-Standing Quotas &amp\; Cryptographic Identity:</strong> Apply ing SPIFFE/SPIRE-based short-lived workload attestation to model inference endpoints\, preventing compromised API keys from allowing unbounded cloud spend.</li>\n<li><strong>Automated Circuit Breakers:</strong> Establishin g sub-minute anomaly tripwires to isolate rogue recursive agent loops and anomalous token velocities in production.</li>\n<li><strong>Return on Auto nomy (RoA) Modeling:</strong> Translating raw token consumption into board room-defensible metrics that balance MTTR reduction against compute overhe ad.</li>\n</ul>\n<h4>Key Attendee Takeaways</h4>\n<ol start="1">\n<li><str ong>Threat Mitigation Blueprint:</strong> Actionable telemetry patterns to identify and neutralize LLMJacking attempts before enterprise token quota s are exhausted.</li>\n<li><strong>Production FinOps Patterns:</strong> Ar chitectural designs for context-window compression and dynamic inference r outing that immediately decrease operational cloud spend.</li>\n<li><stron g>Executive Governance Framework:</strong> Clear policies for chargebacks\ , environment-scoped spend tripwires\, and engineering-versus-token ROI tr ade-offs.</li>\n</ol> END:VEVENT END:VCALENDAR