33,353 results · Most relevant
Issue credentials that are inherently short-lived and issued per task, as opposed to living forever on someone's laptop. Long-lived tokens living forever on someone's laptop create a security risk for agentic AI. Short-lived scoped credentials could have stopped the destructive deletion.
Audit logging must be owned by you, not the agent, and kept outside the agent's reach — an agent that can alter its own logs can rewrite the narrative and fabricate results.
Agent sandboxing acts as a firewall for agents: restrict file system and network access so a compromised agent has no path to credentials and cannot exfiltrate without setting off alarms.
Agent sandboxing can be thought of as a firewall for agents: restrictions on the file system and the network. Execute agents in an isolated remote workspace without a route to production so that regardless of what the prompt says, the agent has no path to do damage. Without sandboxing, a misbehaving or compromised agent has a large blast radius, and agents running wild on people's actual computers and laptops is a dangerous execution environment.
Agents should be executing in isolated remote workspaces; they should never be working on the developer's actual machine and never given a route to production. This is the highest leverage building block on the whole list because it doesn't matter what the prompt says if there's no way to get to production or to the sensitive credentials, you're safe. Hard environmental separation could have stopped the Replit incident: if the agent had been executing in an isolated remote workspace without a route to production, it doesn't matter what the prompt said, the agent would have no path to do damage. It could also have shrunk NX down to the blast radius of just one disposable container — an ephemeral remote workspace with nothing personal on it means the blast radius becomes the one container,
Agents should execute in isolated remote workspaces with nothing personal on them, so the blast radius is one disposable container instead of the whole laptop.
Set up audit logging that you own rather than the agent, so the agent cannot write its own record of its actions. Every agentic action should be attributable to a human being and logged somewhere outside of the agent's own reach. Alert on logs in real time rather than burying them in loads of data — logs buried in data instead of alerted on in real time delay incident detection from minutes to months.
AI tooling should inherit the same identity provider enforcement as human access; it shouldn't be some parallel subscription with a door of its own. Identity-aligned access control is what could have directly stopped the Copilot incident — a finance user's identity should never be able to grant an AI assistant the ability to help someone map out how to bypass the company's data loss prevention measures.
Stick to short-lived and scoped credentials — we are not in a world where persistent broad-scope tokens make sense. Use things like automatic expiration, narrow scoping, and tokens issued per task instead of per issue, rather than credentials living forever on someone's laptop. Short-lived scoped credentials could have prevented Replit, which got its hands on a standing write-access credential that made the deleting possible, as well as the Copilot phishing where the attacker kept using the subscription after the user session ended. In the NX case, short-lived scoped credentials would have meant the attacker could not keep using Copilot after the finance user signed out.
AI tooling should inherit the same identity provider enforcement as human access — identity-aligned access control means a finance user's identity cannot grant an assistant permission to bypass DLP measures.
Credentials should expire on their own and carry only the task's scope — issued per task, not per issue — so destructive actions become impossible and stolen tokens die quickly.
Take steps right now, as soon as this week, to protect yourself and your company from AI incidents, rather than keeping an eye out for things down the line. First, track down every long-lived token that agents in your organization can read. Second, figure out how many of your agents can find access to production, can find some way around your current limits. Third, check where your agents are actually executing — are they sandboxed, in managed workspaces, or just going wild on people's actual computers and laptops? Finally, run a premortem with your team: assume one of the three incidents you just saw happens to you next month. Either ask people what they actually do or pretend the incident is actually happening and see what people actually do. That exercise is going to tell you exactly
A model proxy layer sits between the model (brain) and the agent (body), seeing everything that passes: it can strip injected instructions from external content, flag DLP-bypass requests before they are answered, and filter what the agent sends out.
Dr. Andy Galpin states he prefers roughly three days a week of total body training for most people, because it is more resilient to life disruptions; he notes it is not technically more effective than splits. He also warns that isolating a body part to one day per week risks going about 13 days between workouts for that muscle if a session is missed, which is too infrequent to maintain.
Dr. Andy Galpin recommends increasing training load by roughly 3% to 5% per week as the target rate for progressive overload.
A prompt-driven workflow for generating a type scale with the Figma agent: set a base size, prompt for a type scale with variables, choose a ratio, generate a specimen, and create local styles for each size.
Use the Figma agent via plain language prompts working off the canvas: set a base font size (e.g. BioRhyme 24), prompt it to create a new variable collection with a chosen typeface, request variables for font name, weight, font size, line height, and letter spacing for each size, ask for options for different type scales, choose a type scale ratio when asked (e.g. minor third), prompt it to generate styles for the type scale, name the collection (e.g. 'my agent scale'), and then go in and tweak the values after generation.
Steve Jamison recommends calculating the intelligence cost of a PR by adding what you pay your engineers plus what you pay for AI into a single cost bucket, and tracking that total intelligence cost per PR monthly alongside the AI bill and PR throughput.
The creator recommends that designers build a lightweight design operations guide that maps every type of design deliverable to four things: what the stakeholder gets, why it matters to the business, how long it takes to create, and what you need to do it well. The guide is described as simple to put together, easy to share with stakeholders, and a 'game changer' for designers facing last-minute requests.
Competing on hourly rates makes price comparisons meaningless because one company could take eight hours and another four for the same job. Break the frame by selling by the job instead — offer an estimate rather than an hourly rate — and transition from selling hourly to selling by job exclusively to realtors to escape the race-to-the-bottom dynamic.
Choose sources → to get a library of your own and a daily brief.