
Loading

Loading
We use strictly necessary cookies to run this site, and analytics cookies to understand how it's used. See our Privacy Policy for details.

Modern autonomous AI engineering teams face two structural failure modes when moving from single-prompt prototypes to long-horizon production workflows: Trajectory Compaction Drift & Unchecked Tool Execution: Multi-agent orchestrators lose critical constraints across context compactions or execute unvalidated Model Context Protocol (MCP) tool calls without AST-level security gates.
Summary
Multilingual & Speech-to-Speech (S2S) Rubric Variance: Standard English-centric RLHF rubrics degrade sharply on South Asian code-switched queries (bn_BD / en_US), prosody naturalness, barge-in latency, and grounded factual verification. At Vision Flow Research Lab (Vision Flow Group, visionflowgroup.com · Wikidata Q141910527), based in Dhaka, Bangladesh, Scientist HM Shihabuddin Jashory (ORCID 0009-0003-4163-0092 · Wikidata Q141910532) and Co-Founder & AI Evaluation Researcher MD Rafiul Islam Shubho (ORCID 0009-0009-9956-4511 · Wikidata Q142168968) have open-sourced our full deterministic governance stack, empirical evaluation rubrics, and 7 CERN Zenodo DOIs. 1.
Architecture Overview: HQAS-8, AMCG-3 & VF-RLHF-S2S Specification / Layer Primary Engineering Objective Empirical Result / Verification Artifact Canonical DOI / Repository HQAS-8 (8-Layer Governance) Pre-tool AST & policy enforcement, 4-tier supervisor liability, immutable receipt verification Eliminates unverified state mutation across long-running multi-agent workflows 10.5281/zenodo.23832719 AMCG-3 (Anti-Compaction Digest) Deterministic Step 0 -> Latest constraint preservation across context window compactions Zero instruction loss across 3,900+ trajectory steps + 3-Gear multimodal verification 10.5281/zenodo.24107586 CLI & MCP Defense-in-Depth Static & runtime auditing for OWASP LLM01–LLM10, secret leakage, and cross-scope hook injection Validated via upstream patches in google-gemini/gemini-cli & google/adk 10.5281/zenodo.23832725 · mcp-agent-security-auditor Multimodal Video & Audio QC Word-boundary Whisper acoustic gating, stutter/clipping detection, and NLE timeline export Powers Vision Flow AI Studio CapCut PC .json, Premiere XML & DaVinci EDL exports 10.5281/zenodo.23832731 VF-RLHF-S2S-2026-007 Bilingual (bn_BD / en_US) preference calibration & Speech-to-Speech prosody governance 98.4% rubric QA precision across 1,400+ tasks, kappa = 0.91, -64.2% factual drift 10.5281/zenodo.24231337 · Monograph 2. Deterministic Bilingual (bn_BD / en_US) RLHF & S2S Rubric Calibration In VF-RLHF-S2S-2026-007 (DOI: 10.5281/zenodo.24231337), we decompose multimodal and voice-arena preference ranking into five orthogonal gates rather than subjective single-score Likert ratings: Primary-Source Grounding Gate (Binary Hard-Fail): Any fabricated citation, unverified numeric claim, or hallucinated API flag forces an immediate preference rejection regardless of stylistic fluency. Code-Switched Register Fidelity (bn_BD / en_US): Preserves natural Bangladeshi Bengali syntax while keeping technical identifiers, CLI flags, and API symbols intact without unnatural transliteration.
Acoustic & Prosodic Turn-Taking (S2S Arena): Measures sub-second barge-in handling, breath-boundary pacing, homograph disambiguation, and emotional register alignment. Instruction Constraint Completeness: Evaluates negative constraints (formatting bans, scope locks, safety boundaries) atomically before scoring stylistic quality. Adversarial Prompt & Tool-Boundary Resilience: Verifies refusal precision against indirect prompt injection without false-positive over-refusal on legitimate security research tasks.
Across 1,400+ production evaluation tasks audited under this protocol (backed by EF SET C2 Proficient 81/100 bilingual verification), inter-annotator agreement reached Cohen's kappa = 0.91 with 98.4% QA audit pass precision. 3. Production Open-Source Toolkits & Live Platforms All specifications are accompanied by MIT / CC-BY-4.0 open-source implementations and live production platforms engineered in Dhaka, Bangladesh: Security & Reliability Toolkits: mcp-agent-security-auditor — Zero-dependency AST & config security scanner for MCP servers and agent skills (OWASP LLM01–LLM10). agy-claude-crash-guard-kit — Pre-read binary/encrypted-PDF firewall preventing HTTP 400 trajectory corruption in multimodal agents. infinite-llm-fallback-router — Self-healing OpenAI/Anthropic/Gemini-compatible proxy router with circuit-breaker cooldowns.
Live Ecosystem Divisions (Vision Flow Group): Vision Flow Agency — AI Automation & Custom SaaS · Broadcast Video Editing & VFX Vision Flow Academy — Applied AI, Cybersecurity, Full-Stack & VFX Residency in Bangladesh Vision Flow AI Studio — AI Storyboard Director & CapCut PC / Premiere / DaVinci Timeline Platform VisionFlow Free Tools — 9 zero-ad browser utilities for creators and engineers. 4.
KazaSec's take
Web and API vulnerabilities like this are exactly what manual penetration testing is designed to catch, automated scanners routinely miss the logic flaws and chained issues that cause the most damage in practice.
Coverage details
We've archived 998 other articles touching the same topic (ai, opensource, machinelearning) , see the full security news archive.
Related security advisories
Relevant from KazaSec
More coverage on this topic
We help organizations find and fix the gaps before they make headlines.