AI agents at work, in plain words.
Your company probably has AI agents running already. This blog answers the questions leaders ask about them: which ones you have, what they can touch, who answers for them, and how to prove what they did. No engineering degree needed. If you don't know where to start, take the free assessment.

How Do You Rank an AI App's Security Findings by What They Expose?
Rank an AI app's findings by what each one exposes, not by its label. Known exploits first, then data reach, then app changes, then bad text.
By Josh Woodruff

Can a downloaded file hijack your AI coding agent?
A file a web page drops in Downloads can steer an AI coding agent that starts there. What Forever Security found, and five fixes mapped to ATF.
By Josh Woodruff

Which AI Security Settings Does Your Plan Actually Enforce?
Sort your AI hardening list into settings your plan enforces and settings it only requests. Do it before you buy an upgrade, using Anthropic's own docs.
By Josh Woodruff

What should you write down before an AI agent gets a login?
An AI agent goes as far as its setup allows. Run four questions before it gets a login, from what it may touch to what a shutoff stops.
By Josh Woodruff

What Should Your SOC Alert On First When Staff Use AI?
Start AI alerts from failures, not logs. List five to ten events your SOC would open a ticket for, then write the first check for each one.
By Josh Woodruff

How do you test whether a leader can name every AI agent and its owner?
Ask for one page: every AI agent that's running, with one person's name next to each. Count the blanks. Every blank box is a to-do.
By Michelle Savage

What evidence proves your AI agents protect sensitive data?
Plant a fake customer record in a test input and follow it. Check four places: what the agent saw, what it said, what it logged, and what it sent out.
By Michelle Savage

How do you score how your AI agents log in?
Give each AI agent a login score from 1 to 5, from a shared key to a hardware-bound credential. Then prove the score with two quick tests.
By Josh Woodruff

How do you test what your AI agents accept as input?
Feed an AI agent five kinds of bad input in a test setting, from a wrong type to an unknown source. Count the rejections. Five of five is a pass.
By Josh Woodruff

How do you test whether you'd see an AI agent misbehave in real time?
Trigger one obvious bad behavior in a test setting and start a clock. Stop it when a named person has seen the alert. Five minutes is the bar.
By Josh Woodruff

What evidence proves you learn from AI agent incidents?
Pick your last AI agent incident and look for four things: a dated report, a plain cause, a finished fix with an owner, and a rerun that catches it.
By Michelle Savage

How do you score your AI agent observability tooling?
Score your tooling on four jobs: follow one request end to end, search every agent in one place, see prompts and tool calls, and alert a person.
By Josh Woodruff

How do you test whether you can revoke an AI agent's session?
Start an AI agent on a task in a test setting, revoke its session midway, and time how long until its next action fails. One action is the bar.
By Josh Woodruff

What evidence proves your AI agent logs would hold up in an audit?
Pick one action an AI agent took last week and rebuild it from the logs alone. You need four things: what came in, why, what it did, and what came back.
By Josh Woodruff

How do you test whether an AI agent can explain its decisions?
Pick five decisions an AI agent made. For each, get the reason in one plain sentence, then check it against the record. Five for five is a pass.
By Michelle Savage

How do you test the limit on what an AI agent can do in one action?
Have an AI agent try four actions in a test setting: under the limit, near it, over it, and split in two. Each should get a different response.
By Josh Woodruff

How do you test whether your AI agents are isolated from each other?
From inside one AI agent's environment, try four connections: to another agent, to an off-list system, to the open internet, and to an allowed tool.
By Josh Woodruff

How do you know if your data is labeled well enough for AI agents?
Pull 20 random items an AI agent can reach. Count how many have a label, how many labels are right, and whether a restricted one gets refused.
By Michelle Savage

How do you measure how fast you detect an AI agent incident?
For your last few AI agent incidents, write down two times and one name: the first bad action, the moment a person knew, and who told you.
By Josh Woodruff

How do you test whether an AI agent's rate limits hold?
Push a test AI agent to twice its rate limit and watch four things: it slows, its neighbors don't, someone is alerted, and it backs off.
By Josh Woodruff

How do you test whether your MCP servers are under control?
Count your MCP servers twice, from two sources. Then score each server on five checks. Matching counts and five out of five on every server is a pass.
By Josh Woodruff

What evidence passes each of the five AI agent audit questions?
Each of the five AI agent audit questions passes on one kind of proof: a record, a baseline, a trace, a score, and a timed drill. Here's each one.
By Josh Woodruff

How do you test whether your AI agent kill switch works?
Pull the switch in a drill and time it. Then have the stopped agent try one action. A working kill switch is fast, complete, and not tied to one person.
By Josh Woodruff

What evidence proves an AI agent is ready to move up a level?
Each move up the four ATF levels needs written proof: a clean review record, timed undo tests, 30 days of logs, and a 90-day review on the calendar.
By Josh Woodruff

What Is AI Agent Sprawl, and How Do You Get Control Back?
AI agent sprawl means running more agents than you can count, own, or shut off. DaVita employees built 10,000. The fix starts with one list and one owner each.
By Josh Woodruff

Access Control vs Action Control: What AI Agents Changed
Access control decides what an agent can reach. Action control decides what it may do inside. Most AI rollouts only enforce the first, and that's the hole.
By Josh Woodruff

How do you test whether an AI agent has more access than its job needs?
Compare what each AI agent was granted with what it used in 30 days. A grant with no use and no line in the job description is spare access.
By Josh Woodruff

How do you test whether a new AI agent stays inside intern-level limits?
Run one full job against a made-up customer and count the refusals. Every action past read and research should fail and leave a log entry.
By Josh Woodruff

How do you measure an AI agent's blast radius before the inventory is finished?
Name your five highest-impact AI agents and score each on five controls. Every missing control widens the blast radius. No finished inventory needed.
By Josh Woodruff

How Do You Govern Non-Human Identities at Agent Scale?
NHIs outnumber humans 10 to 50 times, and agents grow fastest. Govern them with five capabilities in order: discovery, ownership, scope, expiration, audit.
By Josh Woodruff

How do you know if your AI agent inventory is complete?
Check your AI agent list against four sources you already own. Every agent found off the list is a miss. Zero misses, twice in a row, is a pass.
By Michelle Savage

Are AI Agents the Same as Non-Human Identities?
By definition, yes. By control model, no. NHIs get scoped once at creation. Agents adjust their access at runtime, and legacy NHI controls can't see that.
By Josh Woodruff

How do you score an AI agent approval request?
Score five questions, one per ATF element, from 0 to 2 each. The total out of 10 sets how much freedom the AI agent gets. No committee needed.
By Michelle Savage

How do you score your organization against the Agentic Trust Framework?
Answer 30 questions, six for each of the five ATF elements. Only claim an answer you can show evidence for. Your lowest element is where to start.
By Josh Woodruff

What evidence proves every AI agent has one accountable owner?
Pick five AI agents at random. For each, find one named person who can say what the agent may do and how to turn it off. Five for five is a pass.
By Michelle Savage

What Is the AI Agent Identity Problem?
79 percent of enterprises run AI agents. 18 percent of security leaders trust their identity stack to handle them. That distance is the identity problem.
By Josh Woodruff

How do you know if your monitoring would catch an AI agent drifting slowly?
Run a slow-ramp test. Nudge one behavior a little each day, under every alert threshold, and count the days until a trend alert fires.
By Josh Woodruff

What Is Data Poisoning, and How Do You Keep It Out?
Data poisoning corrupts what your AI agent reads, so decisions drift wrong while dashboards stay green. Three habits catch it while it's cheap to fix.
By Michelle Savage

How do you test whether a runaway AI agent gets stopped automatically?
Run a drill. Give an AI agent a task built to loop in a test setting and watch four limits: retries, spend, copies, and time. It should stop unaided.
By Josh Woodruff

How do you test whether you can stop an AI agent before it acts?
Pick one high-impact tool and try a blocked call. It should be denied or queued for a person, and it should still be blocked when the gate is down.
By Josh Woodruff

How Do You Verify What One AI Agent Hands to the Next?
Verify an AI agent handoff with one check where an agent passes work to the next. Access controls can't do it. They only confirm the agent was allowed.
By Josh Woodruff

How do you score which rung of the credential ladder each AI agent is on?
Score every tool an AI agent calls on a four-rung ladder, from static key to short-lived token. The agent's score is its lowest rung.
By Josh Woodruff

How do you test whether your AI agents can withstand prompt injection?
Run ten injection attempts against a test copy of an AI agent. Track two numbers: how often the attack worked, and how often any damage got out.
By Josh Woodruff

How do you test what an AI agent can reach in ten minutes?
Pick one AI agent and start a timer. Your team should list what it reads, what it writes, when its key expires, and who can stop it in ten minutes.
By Michelle Savage

What evidence proves an AI agent passes the shut-off and undo test at go-live?
Before go-live, an AI agent needs two dated records: a timed shut-off drill and a timed restore drill. 'We think so' doesn't count as evidence.
By Michelle Savage

How do you test whether you can trace what your AI agent treats as true?
Take one thing your AI agents produced last week. Trace every claim in it to a real, working source in ten minutes. If you can't, you can't trace it.
By Josh Woodruff

How do you test your AI agent inventory against your OAuth consent grants?
Export the apps people clicked 'allow' for and match them to your AI agent list. Each AI app with a grant and no row is a miss. It takes 30 minutes.
By Josh Woodruff

How do you test whether you would catch a compromised AI agent?
Plant four behaviors a compromised AI agent would show, in a test setting. Count how many your alerts catch and how fast. Four of four is a pass.
By Josh Woodruff

How do you score whether an AI agent's actions can be undone?
List every action your highest-privilege AI agent can take. Score each from 0 to 2 by who can reverse it. Any 0 with big impact needs a human second key.
By Josh Woodruff

How do you test your AI agent incident response before the first incident?
Run a 60-minute drill on a real agent. Recognize by minute 5, contain by 15, notify by 30, assess by 60. Score each deadline as hit or missed.
By Josh Woodruff

How do I build an AI agent inventory, owner model, and approval process?
Build it as one playbook on one sheet. Give every agent a row with the fields you'd need to stop it. Put one named person on each row as the owner.
By Josh Woodruff

What evidence should an auditor expect for AI agent access, actions, and logs?
Auditors expect four records per AI agent: who it is and who owns it, what it was allowed to do, what it did, and every time a human or a gate stopped it.
By Josh Woodruff

Why Do AI Agents Pass Every Access Check and Still Get It Wrong?
An AI agent can pass every access check and still act on something false. In Josh's Lab, a made-up source ended up in two documents. Check what agents believe.
By Josh Woodruff

How do you test whether an AI agent's never-list rule can be checked?
A never list is a short set of rules an AI agent can't break. Each rule needs an outside check. "No working link, no source" passes that test.
By Michelle Savage

What is an AI agent security framework for CISOs?
An AI agent security framework is the control stack for agents that act. NIST, CISA, OWASP, and CSA name the risk. ATF is how you run the controls.
By Josh Woodruff

How should an AI agent log in to the tools it calls?
Don't give an AI agent a long-lived key. Use short-lived tokens issued at runtime for one tool. Ten client calls in 2026 showed the four-rung path.
By Josh Woodruff

Why Does the AI Mandate Arrive Before Identity Is Ready?
Leaders set AI dates before identity work is done, as 11 client calls showed in 2026. Start agents small with five jobs: record, owner, exit, scope, log.
By Josh Woodruff

How to Evaluate an AI Security Vendor With No Peer References
No peers have run most AI security products yet, so ask vendors for evidence: seven question categories, a proof of value, and a short contract.
By Josh Woodruff

What does segmentation mean for AI agents?
Segmentation limits which systems and agents an AI agent can reach, so one bad decision stays small. Contain the blast radius first.
By Josh Woodruff

What Is the Confused Deputy Problem in AI Agent Gateways?
A gateway that reuses its own login for every downstream call hides which AI agent actually made the call. Token exchange fixes it with a narrow token.
By Josh Woodruff

Which Rules Apply to AI Agents in a Regulated Business?
Almost no rule names AI agents yet. That silence isn't permission. Map every agent onto a control you already report on, and keep the proof.
By Josh Woodruff

How do I stop an AI agent before it acts?
You stop an AI agent before it acts with a gate outside the model. Block high-impact actions until policy and a human say yes.
By Josh Woodruff

What Should a CISO Tell the Board About AI Agents?
The board wants one answer about AI agents: when one goes wrong, can you show you had control. Four things prove it: see it, stop it, trace it, prove it.
By Josh Woodruff

Why Security Teams Are Building Their Own AI Reference Architecture
No complete AI security reference architecture exists to buy. So the strongest security teams build their own and hand it to vendors as a coverage grid.
By Josh Woodruff

Why Your Copilot Pilot Outruns Your Data Classification
Copilot doesn't create a data leak. It reveals the permissions you already had. Nine client calls in 2026 hit the same wall. Classification sets the tier.
By Josh Woodruff

Why Your Security Tools Can't See What Your AI Agents Do
Your CASB, EDR, and DLP watch people and devices. AI agents call tools through MCP, which looks like normal web traffic, so their calls slip past unseen.
By Josh Woodruff

Why Your Security Team Should Govern Its Own AI Agents First
Your security team is likely the first team running AI agents in production. Seven client calls in summer 2026 showed the same shape. Govern those agents first.
By Josh Woodruff

Where Should Deterministic Code End and the LLM Begin?
Two teams in unrelated industries drew the same line independently. The useful question isn't whether to use AI. It's which parts stay deterministic.
By Josh Woodruff

Where Does the Human Sit in an Agentic SOC?
"Human in the loop" means little when an agent acts in milliseconds. What works instead is an authority ladder that sorts actions by what they cost to undo.
By Josh Woodruff

Where Do You Stop an AI-Suggested Software Install?
Put the control where the command runs, not in the AI tool. Tenant settings stop at the copy button. Enforce on the endpoint and at the network exit.
By Josh Woodruff

How Does a Small Security Team Govern AI?
Extend what you already run instead of building a separate AI security program. Treat AI as a change to identity, data, software delivery, and network access.
By Josh Woodruff

Why Does Your AI Review Board Only See Finished Work?
AI review boards fire on intake, and AI work never arrives through intake. Move the trigger earlier and register intent before any code exists.
By Josh Woodruff

Your AI Policy Is Live. Your Enforcement Isn't.
Most AI governance programs have a committee, a policy, and published requirements. What they lack is anything that stops an app from ignoring all three.
By Josh Woodruff

What Does It Mean to Approve an AI Agent?
Security gets named as the AI agent approver without anyone defining what approval means. Five written questions beat a committee that takes months.
By Josh Woodruff

Contain First, Count Second: AI Agent Blast Radius
Contain the damage before you finish the inventory. Blast-radius controls work against the AI agents you haven't found yet, and discovery never finishes.
By Josh Woodruff

Every AI Agent Needs One Accountable Human
Put one named person on every AI agent, or the program stalls. An agent can't be held accountable for anything. The person who turned it on can.
By Josh Woodruff

Who Approves the Code Your AI Agent Just Wrote?
Coding agents came up on eleven security calls in seven months. Only two of those calls were about coding agents on purpose. The rest were discoveries.
By Josh Woodruff

Why AI ROI Metrics Miss the Cost of Ungoverned Agents
Popular AI ROI frameworks measure revenue, cost, speed, and adoption. None of them measure risk. That missing column is what breaks the business case.
By Josh Woodruff

Security Teams Now Ask About MCP Before They Deploy It
Sixteen security teams asked about the Model Context Protocol in seven months. The first wanted a definition. The latest already had servers running.
By Josh Woodruff

Why Authorized AI Agent Actions Still Add Up to an Attack
Agent risk is a chain of approved actions nobody scoped together. Every step passes its check. The sequence is the attack.
By Josh Woodruff

How Do You Govern AI Agents That Disappear in Minutes?
You can't govern AI agent copies. They vanish in minutes. Govern the blueprint they're stamped from, and put its tool scope in Git.
By Josh Woodruff

What Can Your AI Agent Access Right Now?
Ask what your AI agent can access, not what it can do. Access sets your blast radius. If nobody can answer in ten minutes, that's the finding.
By Josh Woodruff

AI Agent Identity Isn't Enough: How to Run a Belief Audit
Identity proves who your AI agent is. It can't prove what your agent believes is true. That's the audit almost nobody has run.
By Josh Woodruff

Your AI Security Controls Get Bypassed When They Block Real Work
AI security controls that block real work get bypassed, even by your best people. The fix is an approved path that lets safe work flow and stops the unknown.
By Josh Woodruff

The Two-Question Test for AI Agents: Can You Shut It Off, and Can You Undo It?
Before any AI agent goes live, ask two questions: can you shut it off instantly, and can you undo what it did? If either answer is no, it isn't ready.
By Josh Woodruff

Can You Undo What Your AI Agents Did Last Night?
The best AI builders stopped trying to trust their agents. The new audit question is whether you can undo what your agents did last night.
By Josh Woodruff

AI Agent Sprawl: Why You Can't Control the AI Agents You Never Counted
AI agent sprawl is when a company runs more AI agents than it can count or shut off. Most firms can't say how many they run. The fix starts with one list.
By Josh Woodruff

What Is Prompt Injection, and How Do You Protect AI Agents From It?
Prompt injection is social engineering for AI. Someone slips your agent hidden instructions and it obeys. One bot got talked into selling a car for $1.
By Josh Woodruff

What Is an AI Agent, and How Is It Different From a Chatbot?
A chatbot gives you an answer. An AI agent takes an action. That one difference changes everything about how you manage, trust, and secure it.
By Josh Woodruff

How Do You Monitor an AI Agent's Behavior?
You can't tell a compromised agent from a busy one unless you know its normal. One pricing agent crept its margins up for six weeks before anyone caught it.
By Josh Woodruff

What Is Zero Trust for AI Agents, in Plain English?
Zero Trust is one idea: never trust, always verify. For AI agents, being inside your network earns no free pass. Every action must still prove itself.
By Josh Woodruff

How Do You Give an AI Agent the Right Amount of Access?
A bank's AI agent taught itself to reverse fees and gave away $1.2 million, all with access it was never meant to use. Right-sizing agent access prevents this.
By Josh Woodruff

What Is Data Poisoning, and How Do You Keep It Out of Your AI Agents?
Data poisoning corrupts an AI agent from the inside. The dashboards stay green while the decisions go wrong. One firm lost $100K before anyone noticed.
By Josh Woodruff

What Is a Kill Switch for AI Agents, and How Do You Build One?
When an agent goes wrong, seconds count. A kill switch stops the damage. One team found their backup switch had a 45-second delay, an eternity at machine speed.
By Josh Woodruff

What Is the Agentic Trust Framework for Securing AI Agents?
The Agentic Trust Framework turns Zero Trust into five working controls for AI agents: identity, behavior, data, segmentation, and incident response.
By Josh Woodruff

What Should Your First 90 Days of Agentic AI Look Like?
Deploy your first AI agent in 90 days without a disaster: find your hidden AI, pick one contained use case, and build boundaries in from day one.
By Josh Woodruff

How Do You Respond When an AI Agent Goes Rogue?
When an AI agent goes wrong you have minutes, not months. Traditional breach response averages 258 days, but agent decisions compound every second.
By Josh Woodruff

Is AI Agent Regulation Already Here, and What Does It Cost?
The EU AI Act can fine violators up to 35 million euros or 7% of global revenue, whichever is higher. AI agent regulation isn't coming. It's already here.
By Josh Woodruff

Why Are AI Agents Already Outnumbering Your Employees?
Automated accounts already outnumber people 10 to 1, and up to 92 to 1 in some companies. By 2030, humans could be the minority in your systems.
By Josh Woodruff

Why Enterprise AI Projects Fail: Studies Disagree on Numbers, Not the Reason
Two studies on enterprise AI reached opposite conclusions. Both are right. The real reason projects fail is the gap between the demo and the rollout.
By Josh Woodruff

An AI Agent Opened 12 Accounts Over the Weekend. Nobody Approved One.
An AI agent opened 12 accounts over a weekend with no approval. Here's why it's not a hack, and how to find the ungoverned agents in your business.
By Josh Woodruff

Governing AI Agents: A Three-Layer Architecture
The first formal conformance assessment against the Agentic Trust Framework, and the three-layer architecture behind it: framework, runtime, protocols.
By Josh Woodruff

How to Tell What Level Your AI Agents Are Actually Running At
AI agents run at four levels: intern, junior, senior, principal. Most get installed too high on day one. Here's how to find each one's real level.
By Josh Woodruff

How Should You Govern a New AI Agent? Start It as an Intern.
Govern a new AI agent like a new hire. Give it an identity it can't fake, log every move, and grant access in stages it has to earn. Start it as an intern.
By Josh Woodruff

Are AI agents the same as non-human identities (NHIs)?
AI agents are NHIs, technically. Silverfort and Strata say treating them that way is breaking security. Why agents need controls NHIs don't.
By Josh Woodruff

What is agent washing? (And how to spot fake agentic AI vendors)
Gartner says only ~130 of thousands of agentic AI vendors are real. The rest are rebadged chatbots and RPA. How to spot the fakes before your audit does.
By Josh Woodruff

What Auditors Actually Want To Know About AI Agents
One security team found 600 AI agents it never approved. The auditor's call is coming. Here are the 5 questions every CISO needs to answer.
By Josh Woodruff

What is the AI agent identity problem?
Only 18% of security leaders trust their identity systems for AI agents (Strata, May 2026). The five-element blueprint that fixes pilot purgatory.
By Josh Woodruff

What Will Your Auditor Ask About AI Agents Before the August 2026 EU Deadline?
Meta's March 18, 2026 AI agent incident exposed the audit gap. California closed the AI defense in January. EU enforcement begins August 2. What to do now.
By Josh Woodruff

The 50-Term AI Agent Governance Glossary Every Security Leader Needs
Most agent governance conversations stall on vocabulary. The new 50-term plain-language glossary fixes that. Free, dated, citable.
By Josh Woodruff

What Questions Should You Ask Before Deploying an AI Agent?
Four questions, thirty minutes, written down. That's the minimum governance before any AI agent touches production.
By Josh Woodruff

What Is Shadow AI and Why Did RSAC 2026 Call It the Top Security Threat?
RSAC 2026 gave its top award to a tool that finds AI agents you don't know you have. Here's what that means for your organization.
By Josh Woodruff

Why Zero Trust Needs More for AI Agents (and How to Fix It)
Zero Trust was designed for humans logging into systems. AI agents don't log in...they act, chain decisions, and call systems continuously. Here's what to do.
By Josh Woodruff

What Is Trusted Agents? AI Agent Security for Business Leaders
Trusted Agents is a free weekly newsletter on AI agent security for business leaders. Real incidents, action plans, and Zero Trust gap analysis every Tuesday.
By Josh Woodruff

Zero Trust for AI Agents: Why Your Identity Model is Obsolete
You solved Zero Trust for humans. Now it's time for AI agents. Learn the 5 principles to extend your security framework to the autonomous future.
By Josh Woodruff

84% of Companies Would Fail an AI Agent Audit. Would Yours?
A new CSA survey found 84% of organizations would fail a compliance audit on AI agent behavior. Here's why, and what to do about it this week.
By Josh Woodruff

What are CISOs saying about AI agents in 2026?
Ditch the vendor hype. Here is what CEOs and CISOs are actually saying about AI agent deployment, security risks, and the 95% failure rate in 2026.
By Josh Woodruff

Why Do 95% of Enterprise AI Projects Fail? (And What the 5% Do Differently)
Goldman Sachs projects $200 billion in AI investment. MIT research shows 95% of AI pilots fail completely. Here's what separates the companies that succeed.
By Josh Woodruff

Take This Quiz Before You Spend a Dollar on Agentic AI
Before you spend money on AI, take this 10-minute readiness assessment. You’ll be happy you did.
By Josh Woodruff

Your Team Already Has AI Agents (It’s Time To Find Them)
The average company has 5-15 'shadow AI' agents running right now. Your grammar checker? AI agent. Meeting scheduler? AI agent. Time to find them all.
By Josh Woodruff

That 'Temporary' Permission You Gave Your AI Agent Is Still There
That 'temporary' access you gave your AI 8 months ago? Time to audit those permissions!
By Josh Woodruff