35,684 results · Most relevant
Javier Garza recommends that when something changes in your code or AI application, triggers like webhooks should fire a scan; that data pipelines being updated or AI models being deployed should trigger a scan; and that when an agent creates infrastructure-as-code, you monitor that Terraform file so the moment it changes you run a scan.
Javier Garza recommends putting rate limits on AI systems to prevent misuse of infrastructure for unauthorized tasks like crypto mining, limiting exposure and using input validation to defend against embodied consumption attacks, and using resource allocation management so people cannot misuse your AI infrastructure. He also recommends implementing data security and protections for APIs and rate limiting.
Javier Garza recommends that when you run a long prompt and step away, you put verbose logging into a file so a background tool can monitor for unwanted behavior, and that when agents do a lot of things and you don't pay attention, they may do things you don't want, so you should audit them.
Discover and map the AI system structure — for example with an AI-generated map — then take the JSON output into a tool that renders it as a graphical data flow diagram (many tools accept JSON-format maps). Create AI-aware architecture diagrams or use model artifacts, and enumerate threats including adversarial inputs and prompt injection to ensure coverage.
Javier Garza recommends using an existing threat modeling framework (STRIDE, MAESTRO, Attack Trees, DREAD) instead of starting from scratch, and explains that STRIDE covers spoofing, tampering, repudiation, information disclosure, denial of service, and elevation of privilege — cover each of those when you design something.
Dan recommends setting a budget for the agent to avoid surprises like the agent consuming 20K per month in resources.
Dan recommends evaluating the agent's compliance with standards like SOC 2 and internal policies as part of the scorecard.
At runtime, the agent needs to log, create tasks based on the mandate, and execute the tasks. At runtime, verify whether tasks are written, whether the input is logged, and whether the required context is present. Success criteria should be created together with each task, because without success criteria linked to an individual outcome (like a contract renewal), it is impossible to judge whether that renewal was successful. The agent should explain everything it is doing — this is what the task topic is for. At the end, validate the output — there is one single score that evaluates whether the agent was successful or not.
Dan recommends tracking drift against the baseline established at the first go-live date to confirm the agent stays within its boundaries.
Dan recommends using a central action center where people engage with all agents in the organization, rather than each agent handling its own handoff.
Fuad Ali shares an example prompt to run with Claude Code: 'Hey, check my traces for any errors or anything I can fix. Cross-reference with the code for the agent in this repo and use the agent skills.'
Maintain two loops: the detect-investigate-hypothesize-fix loop, and the validation loop that tests the fix on real failures and edge cases from your working example.
endorsed
Use trace data — the source of truth for what your agent is doing — combined with your actual repo data (the code that defines what your agent should be doing) to get really interesting analysis and make improvements at scale. Look at the entire context of what your agent is doing in production, not just individual traces.
endorsed
Frame the agent system prompt as 'you are an expert search agent that uses web search to answer questions accurately, provide valid JSON with the final answer, supporting URLs, and a short justification.'
Run impact analysis pre-flight — not at build time and not in a production incident — so the agent knows the blast radius before it acts on a change.
Yoni Michael lays out his recipe for building solid, honest evals: mine real task candidates out of your corpus, build a deterministic anchored score as much as possible, and unit test the score itself first. Then run evals many times over and track all metrics including pass@k and pass^k, and include examples of things your agent may not be good at — you don't want it to pass 100%, because that's not useful. He adds that evals are your chance to see what your agent will do in as hard as possible circumstances, and that giving agents very easy and simple evals doesn't do anyone any service.
Ensure tools return useful results on the first call every time; otherwise agents that hit dead ends will abandon the tool and return to grep permanently.
To iterate on a function, Dean adds more code — for example, code for text generation — and adds Python dependencies directly in the function code; Flash installs the dependencies on the fly when running the job. Running a Hugging Face model without an HF token produced a warning about communicating with the Hugging Face server, so he enabled his HF token by changing the code to avoid the warning. He then tested text generation with a prompt like 'Who is Batman?', and went for something more intensive than just generating text by testing image generation.
They were trying to get to prevention instead of response, which is why they looked at something wired into the Linux kernel on the front end rather than detect-then-stop. The best way to stop a fork bomb on a Linux box is to set a kernel configuration file that says each process can only spawn 1,024 processes. It doesn't always get fancy — sometimes stopping an attack is just a change on the Linux kernel, and you don't have to run Tetragon or whatever.
You should have Kubernetes network policy, and actually network policy at L4, L7, and wherever you can. Challenge one, the network, is solved by Kubernetes network policy.
Choose sources → to get a library of your own and a daily brief.