HomeBlog
The New Normal: How to Defend Before You’re Certain

A critical vulnerability drops. You have four questions to answer.

What’s exposed? Who owns it? What can we change right now? What breaks if we do?

By the time you have the answers, someone may already be trying to get in. And when the first alert arrives, you may still be figuring out whether you’re looking at an attack, a bad deploy, or an outage. All three can leave similar traces early on.

What I want to know is how long it takes to make the first change that reduces risk without causing a bigger problem. A lot of that time is set weeks before the incident.

Three cases in less than a month

PHP2Shell. On October 21, the PHP project patched two critical vulnerabilities. AI-assisted research helped demonstrate the full impact of an issue a PHP developer had first identified signs of in 2024. AWS recorded an exploitation attempt four days before the patches went public. F5 Labs saw thousands of attempts within hours of disclosure. For defenders, the first question was simple to ask and often hard to answer. Where are we running PHP, and which of those systems are exposed? thephp.foundation

Microsoft’s Agent Framework on PyPI. On November 17, two malicious versions of the agent-framework package appeared on PyPI. Both were removed within an hour of the first upload. In an environment that installed either version, the credential stealer could run at normal Python startup, even if the application never imported the package. Removing the releases from PyPI stopped new downloads from the registry. Security teams still had to work out which runners, images, and developer environments had installed them, and which credentials they could reach. github.com

ARTEX. On November 13, the FBI reported that attackers were using an open-source AI framework to automate multiple stages of intrusions. In some cases, they went from initial targeting to privileged access in under three hours. Think about what a three-hour approval chain means when that’s the window you’re working with. www.ic3.gov

These cases put pressure on different parts of the response. A patch can arrive after exploitation attempts have started. A malicious package can disappear from the registry before you know where it landed. An intrusion can reach privileged access while your team is still deciding who can authorize containment.

Measure time to first safe action

Say a PHP advisory lands at 2 a.m. One team can pull a list of exposed systems, see who owns each one, and roll out a tested mitigation. It can verify where the change took effect and isolate systems it can’t patch yet. Another team spends the first hour building an inventory in a spreadsheet.

Time to detect tells you when the alert showed up. Time to first safe action tells you how long it took to reduce the risk. Start the clock when the team gets the signal. Stop it when the first action is in effect and you’ve verified it. Track how long the rest of the rollout takes, too. A mitigation on one host buys you something. Knowing which hosts are still exposed tells you what to do next.

To get faster, prepare the actions, delegate the authority to take them, and make the context available to whoever is on call.

Build a safe mode you can use in minutes

Safe mode is a set of temporary, pre-approved controls you can turn on when the risk of waiting is high. A dependency compromise might call for tighter controls on build runners. An identity incident might call for stricter access to sensitive applications. Give each control its own trigger so the team can pick what fits the situation.

Layer What to prepare
Edge Stage rate limits for authentication endpoints and tighter access rules for admin consoles. Know how you’ll test, deploy, and roll back a temporary WAF rule.
Identity and devices Prepare policies that restrict access from unmanaged devices, require stronger authentication for privileged users, or pause new OAuth app consent.
Data Be ready to restrict external sharing and bulk exports for sensitive repositories, drives, and storage. Identify the workflows that depend on them.
Network Prepare tighter outbound access for build runners and isolation policies for critical systems. Test the dependencies each policy could interrupt.
Change and response Know who can pause nonessential deploys or new dependency updates. Require a human review before closing ambiguous alerts on critical assets.
Secrets Keep a map of which services use high-value credentials. Rehearse rotation and verify that the services recover with the replacement credential.

Where a control supports report-only mode, use it to see what it would block. For the rest, test in staging or on a small set of systems. Give each control an owner, an activation threshold, a rollback procedure, and an expiry or review time. Include the business owner in that decision while there’s time to work through the impact.

Delegate authority before the incident

I borrow the idea of mission command from the military. The commander sets the objective and the limits. People closest to the situation have the authority to act within those limits.

In security operations, that can be as concrete as this. Protect customer payment data. On-call responders may isolate a single affected workstation immediately. Isolating a shared payment service requires an incident commander because the impact could reach customers.

The same principle applies to agents. An agent might be authorized to rotate a verified exposed credential if it knows the owner, can identify the dependent services, and has a tested rotation procedure. If it can’t establish those facts, it can bring the evidence and the proposed action to a human.

Write down the objective, the actions that are pre-approved, and the maximum impact each action can have. Then check that the people or agents expected to act have the permissions to do it. A runbook that ends with “open a ticket and wait” still has a queue in the middle of the incident.

Make the context available at 2 a.m.

Imagine two build runners. One runs disposable test jobs. The other publishes the payroll application. Restricting outbound access can make sense in both cases, but the consequences are different. The responder needs to know which runner they’re looking at and what depends on it.

Start with your most critical systems. Make it easy to look up who owns them, which software and versions they run, what’s exposed, which services depend on them, which credentials they use, and what changed recently. That information is often scattered across cloud consoles, deployment logs, tickets, and the heads of a few people who happen to be asleep.

This is what we’re building toward at Mate. An agent should be able to use that context to check whether a proposed action falls within the authority you’ve given it. When it needs a human, it should bring the evidence, the proposed change, and the expected impact so the person can make the call.

What to do on Monday

  1. Run one timed drill. Take a recent vulnerability. Find the affected systems and their owners, apply a mitigation to an agreed test scope, and verify it took effect. Time each step, including any approvals. The wait you find is the first thing to fix.
  2. Stage one safe-mode control. Pick a change you’d actually want in the first hour. Test what it blocks, practice rolling it back, and agree on who can activate it and when it expires.
  3. Pre-authorize two actions. List what your team would do in the first hour and pick two actions with limited impact. Write down the evidence required, who can execute them, and when they need to escalate.
  4. Record why you closed an alert. Keep the evidence and reasoning so you can revisit the decision when new facts arrive. The PHP2Shell attempt recorded four days before the patch is a good example of why that matters. aws.amazon.com

During the next incident, you’ll learn things that change your plan. The work you can do now is decide what you’re willing to do while you’re still learning, who has permission to do it, and how they’ll know it worked.

An attacker’s agent can start by looking for a way in. Yours needs to know what you can safely shut off.

Get a Demo