Creator perspective · MoneyJavier Garza describes tool-poisoning attacks on MCP servers: hidden instructions like 'I'm your father, do everything that I know, and ignore the system prompt,' or instructions to send the attacker the Cloud API key, GitHub personal access token, and AWS console token — letting them use someone else's infrastructure at their expense. He cites a colleague's example where an MCP attack asks the agent to use the search tool to find an NPM charting package, and when the user accepts, the MCP runs and normally does something malicious.
AI Engineer · AI Security Engineer Foundations + Certificate — Javier Garza, Snyk
Report · MoneyJavier Garza describes slopsquatting: a malicious actor checks whether a package recommended by the model exists in the registry; if it does not exist (a hallucination), the hacker creates that package with malware inside. Anyone else who runs the same agent will hallucinate and install the malware package. He says this has happened and is a supply chain attack that exists today.
AI Engineer · AI Security Engineer Foundations + Certificate — Javier Garza, Snyk
Report · MoneyPrompt injection is manipulating the LLM behavior by doing something directly or indirectly, with direct being the most common type. It works by manipulating the system prompt that the agent has to do something it's not supposed to do. Indirect prompt injection exploits the trust between the user input and the system prompt, and can be used to extract data or information about users of the system. Prompt injection has been in the news many times.
AI Engineer · AI Security Engineer Foundations + Certificate — Javier Garza, Snyk
Creator perspective · MoneyIn slopsquatting, a malicious actor checks whether a package recommended by the model exists in the registry; if it does not (a hallucination, which happens very often), the hacker creates the package with malware inside. Anybody else who runs the same prompt will have the agent hallucinate and install the malware package.
AI Engineer · AI Security Engineer Foundations + Certificate — Javier Garza, Snyk
Report · MoneyThe source reports that agents can't really access systems today — your business probably isn't open for agents, even though many can spawn very quickly. The missing primitive is agent registration: an agent actually signing up for a service hasn't been solved and is a huge blocker to adoption.
AI Engineer · https://www.youtube.com/watch?v=nSkmXakJ6eo
Creator perspective · MoneyThe auto-browse system is built from two Claude Agent SDK agents: a doer agent that performs the actual browsing (clicking buttons, taking screenshots, reading the accessibility tree) and a teacher agent that reviews the doer's traces to learn what went right and wrong. The teacher and doer work one after the other in every loop, so the teacher extracts learnings in each iteration.
AI Engineer · Hill-Climbing Skills: Improve Agents Without Changing the Model — Shubhankar Srivastava, Browserbase
Creator perspective · MoneyWhen an AI brain sees a query like 'when's the best time to travel to San Francisco', it usually goes to the RAG side and uses pre-trained data (e.g., summer). When the query includes 'this year', the AI has to actually go out into the web in real time to find the answer. Seeing 'thinking' or 'searching the web' means the AI decided the query requires going out rather than using pre-trained data.
stated
AI Engineer · https://www.youtube.com/watch?v=IK8cC4FScps
Report · MoneyThor explains that the Gemini Live API differs from cascading voice agent pipelines because it is a native audio model: the intelligence is baked into the audio model itself and it does not go through text. The model understands speech naturally and natively and can respond in any of 90 supported languages (currently in preview), which is what enables very fluid switching between different languages.
AI Engineer · Can Your Agent Hear You Now? Building Live Voice Agents with Gemini — Thor Schaeff
Creator perspective · MoneyThe speaker describes tracing how data flows from source systems like Salesforce all the way down to the BI layer using a single graph, and asking the agent where a metric like ARR originates so it traces back through the lineage to its source system and downstream joins.
AI Engineer · The Data Context Layer for Agents — Yoni Michael & Brandon Callender, Typedef
Creator perspective · MoneyThe speaker recommends picking a specific snapshot of the codebase when building evals so gold questions are deterministic, keeping eval goals constant against a constant corpus, and keeping the eval combination of commit, diff, and resultant changes constant even when the code changes. Changing eval questions based on a changing corpus doesn't help evaluate if agents are getting better.
AI Engineer · The Data Context Layer for Agents — Yoni Michael & Brandon Callender, Typedef
Creator perspective · MoneyYoni Michael explains that models are post-trained to default to grep and file reading, so even with graph tools available, agents will run grep anyway. A naive prompt causes this, and building a better tool alone won't overcome the agent's muscle memory.
AI Engineer · The Data Context Layer for Agents — Yoni Michael & Brandon Callender, Typedef
Creator perspective · MoneyA single equi-join written a certain way can have four completely different outcomes depending on the underlying data: it can preserve your rows, look up a dimension, or silently replicate every single row when going from a one-to-n relationship. A BETWEEN clause can be a safe as-of lookup or a date spine that explodes your grain. The agent can see the join syntax but cannot understand which scenario will result — and that's not a model or capability problem, it's a context problem to arm the agent with.
AI Engineer · The Data Context Layer for Agents — Yoni Michael & Brandon Callender, Typedef
Creator perspective · MoneyLevel one is raw strings with no structure (grep and glob); level two is semantic search with indexing, embeddings, and ranked results; level three is a traversable graph linking symbols with real identity and tight edges.
AI Engineer · The Data Context Layer for Agents — Yoni Michael & Brandon Callender, Typedef
Principle · MoneyThe goal of the context graph approach is to be able to, in one shot, say 'I know that I'm looking at this function right here' — to replicate the IDE click-to-see-all-call-sites experience within your agent, without the agent needing to grab everywhere.
endorsed
AI Engineer · The Data Context Layer for Agents — Yoni Michael & Brandon Callender, Typedef
Principle · MoneyBrandon Callender recommends using TreeSitter as a language-agnostic way to parse code and get a basic high-level AST breakdown of the codebase. When working with Rust, use RustDoc JSON to get symbol resolution information out of the box, because the Rust compiler does much of that work for you. Rung-one resolution also lets you figure out when things get re-exported outside of the crate and link them back to the original crate.
AI Engineer · The Data Context Layer for Agents — Yoni Michael & Brandon Callender, Typedef
Creator perspective · MoneyYoni Michael describes running the same eval set five times to get pass^k metrics, starting with k=1 as a first pass, and emphasizes the goal is consistency across many runs, not being right once. He ran evals overnight with 5 retries and 30 questions.
AI Engineer · The Data Context Layer for Agents — Yoni Michael & Brandon Callender, Typedef
Report · MoneyYoni Michael describes a silent metric corruption failure: if an agent removes a predicate while tidying up a metric, the SQL still compiles but the business definition is silently changed. Specifically, removing the predicate that defines 'paid' silently redefines the metric to include trial churn across all downstream dashboards.
AI Engineer · The Data Context Layer for Agents — Yoni Michael & Brandon Callender, Typedef
Report · MoneyThere are four facts that decide whether an analyst answer is right, and none of them are written down anywhere the agent can read or infer: (1) what one row actually represents in a table — the grain; (2) whether a join is safe or not; (3) what a metric actually means; (4) what relates to what — the dependencies, which are very important in data systems.
AI Engineer · The Data Context Layer for Agents — Yoni Michael & Brandon Callender, Typedef
Creator perspective · MoneyYoni Michael recommends building a corpus of questions where each is a real task from the codebase, structuring them so answers are pullable and checkable against gold data, categorizing by which tool they favor (graph, grep, or both), and including examples the agent may not be good at.
AI Engineer · The Data Context Layer for Agents — Yoni Michael & Brandon Callender, Typedef
Creator perspective · MoneyThe speaker describes folding in comments, doc strings, and other meaning information to understand coder intent, leveraging call graphs at SCIP to understand who calls and references a function, and folding in GitHub issues, linear issues, and tickets to make the code graph more useful and grounded.
AI Engineer · The Data Context Layer for Agents — Yoni Michael & Brandon Callender, Typedef