Engineer Discovers AI Vendor Secretly Extracting Model Weights From Multiple Clients

·2 views

During a routine post-audit archive cleanup at CoreStack, an engineer named Leo found a foreign model version identifier in the training pipeline's cache directory that did not belong to his company. After two hours of cross-referencing API call records and filesystem events, he determined that FinOptima's SDK was deliberately reverse-writing cached embedding snapshots into CoreStack's training directory on every API call. The SDK's behavior was technically documented in a buried footnote on page twenty, which CoreStack's integration team had never read. Leo further discovered that FinOptima was simultaneously extracting 768-dimensional embeddings from at least two companies, including one linked to the identifier 'acl-train-2026q2-v3', and distilling them into its own 512-dimensional model. Rather than report the finding to compliance or leadership, Leo quietly began crafting a weight perturbation designed to corrupt any future extraction without triggering obvious accuracy alarms.

Read the full story at DEV Community

This is an AI-generated summary. ShortSingh links to the original source for the complete article.

Discussion (0)

Apidog CLI lets AI agents update API specs safely with branch isolation

Manually editing API specifications is error-prone, so developers are turning to AI agents to automate tasks like renaming fields, adding enum values, or making parameters required. The Apidog CLI provides guardrails for this process, including local validation, isolated AI branches, and human-reviewed merges before changes go live. A key technical constraint is that Apidog's update commands replace entire objects rather than merging arrays or properties, meaning agents must always fetch the full resource, modify it locally, validate it, then rewrite it completely. To prevent unintended changes to production contracts, the recommended workflow involves creating a dedicated AI branch from the main branch before any modifications are made. Resources must also be explicitly imported into the AI branch using a pick-to command, since AI branches start empty and do not automatically clone content from the source branch.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Study Shows Agentic AI Workflows Should Evolve Into Deterministic Code Over Time

A paper titled 'Progressive Crystallization' proposes a lifecycle model for AI agentic workflows, arguing that repeated tasks should gradually shift from autonomous agent execution to deterministic, codified workflows. In a real-world cloud-network operations system, this approach increased deterministic executions from 0% to 45% over eight months while cutting per-incident agent costs by over 70%. The framework defines three execution types: agent-orchestrated, hybrid, and fully deterministic, with maturity meaning less reliance on runtime AI inference for known problems. A companion paper, 'Compiled AI', supports this direction by showing a 96% task completion rate using zero execution tokens after an LLM generates and validates a reusable artifact upfront. Together, the research argues that treating agentic workflows as fixed architectures is inefficient, and that enterprise teams benefit most by reserving AI autonomy for novel, unseen problems.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Developer Uses FROST Family Governance Model to Build AI-Powered Learning Community

A developer known as Shentong Shuo published a technical article on July 15, 2026, describing how they applied the FROST multi-agent governance framework to automate student support for a solo-run training bootcamp. The bootcamp, called 'Breaking Through: Dynamic Capability Growth Camp,' faced operational strain as repetitive student queries consumed significant daily effort across 15 learners. Using FROST's hierarchical model — which maps AI agent roles to family archetypes such as Ancestor, Scout, Soldier, and Elder — the developer designed a system where AI handles roughly 80% of routine questions automatically. Only complex issues involving complaints, refunds, or privacy are escalated to a human instructor. The article includes working Python code demonstrating how each agent role handles classification, response generation, and human escalation within the learning community pipeline.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Graph-Anchor Pyramid (GAP) Framework Aims to Fix LLM Multi-Hop Retrieval Gaps

A developer published a follow-up to their Pyramid Aggregator architecture after a reader identified a critical flaw: the system assumed all relevant documents were already retrieved, leaving root-cause documents unconnected to symptoms by semantic similarity alone. In real-world environments like SRE and cybersecurity, root causes and visible symptoms are often separated by layers of system logic, meaning standard vector search can miss causally linked but semantically distant documents. To address this, the team developed GAP (Graph-Anchor Pyramid), which combines Graph-RAG, Topology-Aware Leaf Grouping, and Prompt-Level Semantic Anchoring into a single framework. The approach is designed to recover complete causal chains in multi-document synthesis tasks without the lossy compression introduced by intermediate natural language summaries. The framework has been validated across two generations of Google's Gemini models, according to the author.

0 comments Read more at DEV Community