When to Update Firmware (and When to Leave It Alone)
Every firmware decision sits between two proverbs, and both of them are right. The first is "if it works, leave it alone", because a working network is a precious thing and updates sometimes break it. The second is "unpatched is unsafe", because every old firmware image carries every hole ever found in it. Attackers read the same advisories you do. Ask two experienced admins which proverb wins and you will get three opinions, each delivered with conviction.
The way out of the argument is to stop having it. The answer is not a mood. It is a policy: the same few questions asked of every release. The answers route each update into an expedite lane, a scheduled cycle, or a documented "no". Shops that struggle with firmware decide each update emotionally. They feel fear after the last regression and panic after the last frightening headline. They end up with the worst of both: critical fixes delayed and risky updates rushed. I ran a network that way for a while. It is tiring, and it is not safer.
This guide gives you the policy. It covers the four questions to ask of any firmware release. It offers a fair look at the "stay one version behind" philosophy, plus a review cadence per device class. It also covers the habit that makes skipping an update a decision instead of an accident. It pairs with the firmware inventory, which tells you what you run today. It also pairs with the safe upgrade procedure, which covers how to apply what you decide.
In this guide
- The tension, stated honestly
- Question 1: What does it actually fix?
- Question 2: How urgent is it?
- Question 3: What could it break?
- Question 4: When should it happen?
- The n-1 philosophy, treated fairly
- A review cadence that runs itself
- Skipping with your eyes open
- The whole decision on one page
The tension, stated honestly
Both proverbs earned their scars. Firmware updates genuinely do fix security holes and real bugs. And they genuinely do introduce regressions: new defects in things that used to work. Network gear regressions are especially memorable because the blast radius is everyone. Ask around and you will collect the stories. They include the update that quietly broke a VPN option one site depended on. There is the release that made a switch forget a setting on reboot. There is also the new version whose "improvement" changed a default and took an hour of a maintenance window to hunt down. Every admin who has lived one of these carries a small, reasonable flinch.
But the other failure mode is quieter and more dangerous. A box that nobody updates does not stay the same — it rots. The firmware is unchanged, but the world around it keeps discovering holes in it, publishing them, and building tools to exploit them. A firewall running a years-old release is carrying every vulnerability found in that release since the day it shipped. They are all documented in public, all fixed in versions you chose not to install. Nothing about that shows up on a dashboard. The network hums along, green lights everywhere, while its exposure grows one advisory at a time.
Notice the asymmetry in how the two failure modes are remembered. The regression is a story with a date, a villain, and an audience. It gets retold, and it trains the whole team to fear updates. The rot never gets a story, because the day a vulnerability was published in your firmware, nothing visibly happened. Human memory keeps score of exactly one side. That is why gut feel drifts toward "leave it alone" even in shops where the math says otherwise.
So the honest position is: updating carries a small, immediate, visible risk; not updating carries a growing, invisible one. A policy exists to price both risks deliberately, per release, instead of letting whichever proverb shouted last make the call. The four questions below are that policy.
Question 1: What does it actually fix?
Every decision starts with the release notes — the vendor's document describing what changed in the new version. Reading release notes properly means neither skipping them nor treating every line as urgent. You are doing a triage read, sorting the contents into three piles:
- Security fixes that touch features you use or expose. These are the pile that matters most. A fix for the web management interface matters if your web management interface is enabled. A fix for a VPN protocol matters if you terminate that VPN.
- Bug fixes for symptoms you actually have. If the notes describe your exact ghost — the memory leak that matches your reboot mystery, the crash signature you have seen — the update graduates from optional to desirable.
- Everything else. New features you will not use, fixes for hardware you do not own, improvements to protocols you have disabled. This pile is allowed to be large, and it is allowed to bore you.
Some release notes say, in full, "various improvements and bug fixes". Those have sorted themselves: three piles, contents unknown. When the vendor does name things, security items usually arrive with a CVE identifier. That is an entry in the public Common Vulnerabilities and Exposures catalog. The catalog gives every disclosed vulnerability a unique name so vendors and admins can talk about the same hole. The common mistake is treating every CVE as a fire. The useful question is relevance: is the vulnerable service even enabled on our box, and can anything untrusted reach it? A critical-sounding vulnerability in a management interface that is reachable only from your locked-down management network is a real issue with a modest fuse. The same vulnerability on a box whose management page faces the internet is a drop-everything problem. Your inventory answers "do we run the affected version". Your own configuration answers "is the hole reachable"; only both together tell you what the fix is worth.
Tip: read the known issues section of the release notes before the fixes section. Vendors list the bugs still present in the new version. Finding your must-have feature on that list is the cheapest regression you will ever avoid.
Question 2: How urgent is it?
Once you know what a release fixes, sort it into one of two lanes. The expedite lane is for security fixes where the vulnerability is being actively exploited in the wild. It also covers security fixes where the affected service on your device is exposed to untrusted networks. These jump the queue: you schedule a window measured in days, not months, because the race is real. Vendor advisories usually say when exploitation has been observed — that phrase is your tripwire.
The scheduled lane is everything else, and it should be most things. A stability fix for a bug you are not hitting can wait for the normal cycle with no guilt whatsoever. A security fix for an unreachable service rides the cycle too. It still goes in, because configurations change and defense in depth is cheap. But it goes in calmly. The whole point of having two lanes is that the existence of the expedite lane lets you be relaxed about everything in the other one. Shops without the distinction treat everything as either an emergency or ignorable, and both halves of that are wrong.
A pair of examples makes the sorting concrete. An advisory says the remote-access service on your internet-facing firewall can be crashed by anyone who can reach it. In this example, exploitation has been observed. That means the expedite lane, a window this week, no debate. The same week, the switch vendor ships a release fixing a memory leak that appears only when a feature you have never enabled is turned on. That means the scheduled lane, the next quarterly window. Sleep well. Same word — "update available" — two entirely different clocks.
Question 3: What could it break?
Urgency argues for updating; this question prices the other side. Three factors set the risk of any given update.
The track record of this vendor and train. Your own history is data: has this product line's maintenance releases been drama-free for you, or does every second update come with a surprise? The vendor's own release notes are data too. A long parade of fixed regressions in recent releases tells you how much faith the current one deserves. Some product lines have earned the right to be updated casually; others have earned a canary and a long soak. Grade them individually.
How critical the device is. A regression on a lab switch is a story; the same regression on the core switch is an outage. The blast radius of the box scales how much caution the update deserves. That is exactly why the safe-upgrade procedure has you try updates on a low-stakes representative device first.
How easy the road back is. Gear that keeps two firmware images and can boot the previous one makes rollback a five-minute decision. That makes updating dramatically less scary. I have made that five-minute decision at three in the morning and been grateful for every one of the five. A device where downgrading is awkward or config conversion is one-way deserves more testing before you commit. Rollback mechanics — and everything else about executing the update without drama — are the subject of the safe firmware upgrade procedure. For the decision, it is enough to know which kind of device you are dealing with.
Question 4: When should it happen?
For the expedite lane, the answer is "the first responsible window you can arrange." For everything else, the answer is the cycle: a standing rhythm of planned maintenance windows into which routine updates are batched. Instead of fifteen separate decisions about when to update fifteen boxes, you decide once — say, one evening per quarter per site. The scheduled lane simply queues up work for the next window. Batching also means one announcement, one change record, and one set of verification checks instead of fifteen. How to schedule, announce, and suppress monitoring for those windows properly is covered in our guide to planned downtime.
The cycle does something psychological, too: it makes updating normal. A shop that updates only when frightened does it rarely, under pressure, with rusty procedures. A shop with a cycle gets regular practice, so each individual update is smaller, better-rehearsed, and less likely to surprise anyone. Boring is a skill, and it is built on repetition.
The n-1 philosophy, treated fairly
Sooner or later someone will tell you their policy is n-1: never run the newest release; stay one version behind. The reasoning is sound as far as it goes. The newest release has been tested by the vendor. The release before it has been tested by the vendor and thousands of other people's networks. Waiting lets someone else's fleet find the regression, the vendor patch it, and you skip the whole episode. For stability-critical gear, deliberately trailing the bleeding edge is a respectable posture, not cowardice.
But n-1 fails exactly where it matters most: when the security fix you need exists only in the newest release. Refusing version n because of the policy means running a known hole to avoid a hypothetical regression — a bad trade whenever the hole is reachable. So an n-1 policy is only honest with an exception lane written into it. Security content relevant to us overrides the waiting period. In practice, mature vendors soften the dilemma by back-porting security fixes into older releases. When they do, n-1 and patched can both be true. When they do not, patched wins.
Vendors organize releases into parallel trains (also called branches). There is a maintenance train that receives only bug and security fixes. There is also a feature train where new capabilities land along with the new bugs that escort them. For infrastructure you want boring, ride the maintenance train of a mature release. Apply n-1 thinking within it. Let a fresh maintenance release soak for a few weeks while you watch the vendor's forums and defect list. Then take it in the next cycle. That combination — stable train, short deliberate lag, security exception lane — is what most experienced shops actually run, whatever they call it.
Warning: watch for "n-1" quietly decaying into "n-never." If your inventory shows gear five releases behind, the policy stopped being a stability posture a long time ago. In that case, it became plain neglect with a respectable name. The drift view in your inventory is the audit; the cadence below is the cure.
A review cadence that runs itself
None of the questions above get asked unless something puts releases in front of you. That something has two parts. The first is the advisory subscription habit. Every vendor whose gear you run has a security-notification mailing list or feed. Subscribe to each one and route them to a folder you actually read. Then advisory awareness stops depending on luck or trade-press headlines. (The same subscription habit, pointed at end-of-life announcements, feeds your EOL tracking — one habit, two registers.)
The second part is a standing review rhythm per device class. It is a calendar slot where you check what releases exist for each fleet and run the four questions. "Review" does not mean "update"; most reviews end with "nothing worth taking, see you next quarter," and that is a successful review. Sensible starting points:
| Device class | Review rhythm | Reasoning |
|---|---|---|
| Internet-facing gear (firewall, VPN, anything exposed) | Monthly, plus advisories on arrival | Largest attack surface; security content dominates; expedite lane lives here |
| Core switches and routers | Quarterly | Stability outweighs novelty; few exposed services; regressions expensive |
| Access switches and APs | Quarterly to twice a year | Many boxes, modest stakes each; batch into fleet waves |
| Server management controllers (BMCs) | Quarterly, alongside server patching | Security-sensitive and chronically forgotten; piggyback on an existing rhythm |
| UPS cards, printers, slow-moving appliances | Annually, plus advisories | Releases are rare; update when a real fix appears, not by reflex |
When an advisory does arrive between reviews, give it a ten-minute triage on arrival rather than letting it marinate. Open the inventory. Check whether any of your gear runs an affected version. Check whether the vulnerable service is enabled and reachable. Either open an expedite ticket or file the advisory against the next review with a one-line note. Ten minutes is cheap; the alternative — rediscovering the advisory weeks later and wondering how long you were exposed — is not.
Tune the rhythms to your reality — the principle is simply that exposure sets frequency. Gear that talks to the hostile world gets reviewed often; the stable backwater gets a yearly look and an advisory tripwire. Write the cadence down, put the reviews on the calendar, and the policy now runs on rails instead of memory.
Skipping with your eyes open
The step almost everyone omits: declining an update is a decision, and decisions get written down. When a review ends with "skip," record it. Include the device or fleet, version offered, the decision, the reason, and the trigger that would reopen it. The record can be one line in a small ledger kept next to the firmware register:
| Fleet / train | Version offered | Decision | Reason | Revisit when |
|---|---|---|---|---|
| fw-edge / 7.1 | 7.1.4 | Skipped | Only fixes a VPN mode we do not use; no security content | Next monthly review |
| core-sw / 4.2 | 4.2.9 | Scheduled | Advisory fix for management interface | Next maintenance window |
| ap-fleet / 6.0 | 6.0.2 | Deferred | Known-issues list includes a roaming regression | When the follow-up release lands |
The ledger changes the character of skipping. Without it, "we didn't update" is indistinguishable from "we never looked." Six months later even you cannot remember whether the old version is a considered choice or an oversight. With it, every version you run is either current or explained. That is exactly the standard an auditor, an insurer, or your own future self will hold you to. It also quietly fixes the drift problem: a revisit trigger written down is a revisit that happens.
At Northgate Retail the edge firewall ran a release that worked. And "it works" had been the policy for as long as anyone could remember. An advisory for its remote-access service arrived in spring and was filed, unread, under a rule that said updates happened in the autumn window. The notice that exploitation had been observed arrived in summer. The window moved to a Tuesday at lunchtime, which is not a window. The firewall was updated with the whole company watching the login page. Nobody had decided to run the hole for four months; nobody had decided anything. The ledger was started the following week, and its first entry was the one that should have been written in spring.
The whole decision on one page
Everything above compresses into a walk you can run on any release in about five minutes, release notes in hand:
| Step | Ask | If yes | If no |
|---|---|---|---|
| 1 | Does it fix a security hole in a service this device runs or exposes? | Go to 2 | Go to 4 |
| 2 | Actively exploited, or reachable from untrusted networks? | Expedite lane: window within days | Go to 3 |
| 3 | — | Scheduled lane: queue for the next cycle window | |
| 4 | Does it fix a bug we are actually experiencing? | Scheduled lane; sooner if the pain is real | Go to 5 |
| 5 | Does it add a capability we concretely need now? | Treat as a small project: canary and test first | Go to 6 |
| 6 | — | Skip with eyes open: ledger entry, reason, revisit trigger | |
Run every release through the walk and notice what disappears: the mood. No more updating out of vague dread, no more skipping out of vague fear. Each version you run becomes the output of a repeatable decision. And when someone asks "why is this box on that version?", the answer exists on paper instead of in somebody's stomach.
Where to go next
Both proverbs are still right, and you will never settle which one wins. The policy is how you stop needing to. It gives you four questions, two lanes, a cycle on the calendar, and a ledger for every release you decline. Each version you run is then either current or explained. And the flinch after the last regression stops being the thing that sets the schedule. Ask the same two admins a year later and you will get one answer, written down, which is a first.
The policy needs two companions. The firmware inventory tells you what you run and which advisories touch you. And the safe upgrade procedure turns "yes, update" into a boring evening instead of an adventure. Both live in Firmware & Lifecycle Tracking. Alongside them is the longer-horizon question of when a device stops getting updates at all, and what to do before that day arrives.
Put this into practice with SysPeep Monitor
Update decisions lean on knowing your gear: what version each box runs, how it behaves before a change, and whether anything degraded after one. SysPeep Monitor watches your devices over SNMP, WMI, and SSH and alerts you when health or performance shifts. So when you do take an update, you will know quickly whether the network agrees it was a good idea. The free trial runs on any Windows machine.
Download Free TrialLearn About SysPeep