A scoff in lab
SSH keys make me uneasy, and I’ve stopped pretending they don’t.
Recently in lab I made the case. I said “if your machine is compromised,” and my professor, whom I genuinely love, gave a playful scoff. The vibe in the room was clear: that’s an edge case, not a threat model.
I manage IT for a living. I’ve seen remote access trojans on real machines in a real business. So I want to lay out why that phrase isn’t an edge case to me, and why the ease of ssh prod frightens me more than it comforts me.
What everyone gets right
Let me get this out of the way: I am not arguing for passwords.
On the question SSH keys were built to answer, they win easily. Can someone on the internet guess, brute-force, phish or credential-stuff their way into my server? With password login disabled, no. That’s real, and it’s why “use keys, not passwords” is good advice.
But that’s the only question most people ever ask about keys. “SSH keys are secure” is a statement about authentication strength. It says nothing about what happens when the machine holding the key is no longer yours.
To be fair to my professor, a passphrase encrypts the key at rest, so a stolen key file alone isn’t enough. In practice, though, keys are rarely at rest. Plenty are generated with no passphrase. Others sit decrypted in ssh-agent all day, or unlock automatically from the macOS keychain, for example. Encryption at rest only helps if the key is actually resting.
One machine is two machines
When I type ssh fedora and get an instant shell on a production box, that convenience has a precise meaning. My laptop is trusted by that server. Anything that can run as me on my laptop can do what I just did.
An SSH key isn’t really a lock. It’s a written-down trust relationship. Putting machine A’s public key in machine B’s authorized_keys doesn’t say “B trusts me.” It says “B trusts whoever controls A.” Most of the time those are the same person. The whole problem lives in the gap between them.
Security people call this transitive trust: if B trusts A, then anyone who owns A effectively owns B. Nobody decides that. It just follows. So I think of it this way: if machine A has a key to machine B, then A and B are one machine with two addresses. Compromise one and you’ve compromised both.
And it isn’t only SSH. Cloud CLI tokens in your home folder, logged-in browser sessions, an unlocked password manager, a saved VPN profile. Anything that lets your machine act without asking you first extends the same trust. Try listing everything your daily laptop can reach without you lifting a finger. For most of us it’s a sobering list.
I’ve watched this play out
I run IT, infrastructure and security for a multi-location business. I’ve found remote access trojans on machines in that environment. These were real infections on real employee computers, not lab exercises.
Here’s what kept those incidents from becoming disasters. The servers those users normally remoted into weren’t one command away. Connecting took deliberate steps, and it required a second factor the malware didn’t have. The attacker owned the endpoint and still hit a wall.
Now picture the same infection on a machine where getting into the servers was as easy as ssh {alias}. The attacker wouldn’t need an exploit, a phishing email or a password. They’d just need to wait for a shell, then type what the user types. The friction everyone complains about was the only thing between “one infected PC” and “we lost the servers.”
The biggest theft of the year started on one laptop
If this still sounds hypothetical, look at Binance. On November 3, 2026, the exchange published its initial findings on the roughly $6.3 billion stolen from its hot wallets on October 29. The FBI has tied it to North Korea’s TraderTraitor group. Here’s the chain as reported:
- June: An engineer goes through a fake job interview. The “coding assignment” hides malware that replaces the pyenv executable with a backdoored version.
- The following months: The attackers sit on that machine and work their way through internal systems.
- October 20: They use Binance’s own internal AI tool to take control of wallets tied to the first engineer and to a third employee.
- October 29: The wallets are drained.
No one broke Binance’s cryptography. Nobody brute-forced anything. One trusted laptop fell, and the tooling that trusted that laptop did the rest. As one write-up put it, internal automation becomes an access multiplier for anyone who already has a foothold.
Swap “internal AI tool” for ssh prod and it’s the same story. “If your machine is compromised” wasn’t an edge case here. It was step one.
Endpoint compromise isn’t rare anymore
Binance is the headline, but the background numbers make the same point.
- Infostealers are an industry. Flashpoint counted 7.4 million devices infected with infostealer malware in the first half of 2026, up 27% from the previous six months, and about 1.7 billion credentials harvested.
- They reach most companies. Outpost24 found infostealer-related credential exposure at 72% of the 4,824 organizations it studied. In October, attackers reportedly used one Meta employee’s stolen credentials to reach some of the company’s GitHub repositories.
- They get past MFA. Flare found about 1.17 million infostealer logs that held both credentials and live session cookies. That’s enough to replay a session and skip MFA entirely.
- They’re coming for the people with the keys. Doctor Web reported more than 2,000 PCs infected through security-themed GitHub projects that dropped AsyncRAT and infostealers, aimed at security professionals. Separately, attackers probing exposed Flowise instances went looking specifically for SSH keys.
- The clock is short. Palo Alto Networks reported in October that 29% of newly disclosed CVEs are exploited within 24 hours.
Assuming your laptop stays clean is becoming increasingly optimistic. Security means planning for compromise. Not inviting it, but being prepared if it should happen. Defense in depth.
If you need the access anyway
Most of us do need remote access to production. The goal isn’t to give it up. It’s to make sure a compromised laptop doesn’t quietly become a compromised server.
- Gate the network path. The gold standard is Tailscale or another WireGuard-based VPN. Prod stops answering the public internet entirely, and only devices you’ve enrolled can reach it. Tighten it with ACLs so only chosen devices can reach SSH at all.
- EDR on both ends. It should be obvious, but run a trusted EDR on prod and on the machine you remote in from. If something lands on your laptop, you want to know before it reaches the server.
- Hardware-backed keys. Use an
ed25519-skkey on a YubiKey. The private key never leaves the device, and every connection needs a physical touch. Malware can’t SSH to prod at 3 a.m. on its own.
None of these are exotic. They’re the standard answers to the exact problem I’m describing.
The rule
Here’s my position, and I’ll defend it all day.
If you’re comfortable with one machine being two machines, go ahead and use a plain key on disk. If you’re fine with the idea that anyone who gets onto machine A also gets onto machine B, that’s a legitimate choice. Make it on purpose.
If you can’t accept that model, don’t build it.
“SSH keys are secure” answers whether someone can guess their way in. It doesn’t answer what falls when your laptop falls. I think that second question deserves a lot more airtime than it gets.
Sources
- Binance says hackers used fake job interviews and AI tools to steal $6.3 billion (Tech Startups)
- Fake job interviews: how $6.3 billion was lost from Binance and Binance: hackers stole $6.3 billion with fake interviews and AI abuse (SecNews)
- MaxwellCTI Weekly, Oct 27 to Nov 2, 2026 (FBI attribution to TraderTraitor)
- Infostealers harvest 1.7 billion credentials in six months (Infosecurity Magazine)
- How infostealer logs became the fuel (Cyber Security News, citing Flare)
- Infostealers expose 72% of organizations (eSecurity Planet, citing Outpost24)
- The Hacker News weekly recap, November 2026 (Doctor Web GitHub campaign, Flowise SSH key hunting)
- Critical PHP vulnerabilities: Frontier virtual patching customers were protected before day one (Palo Alto Networks, Oct 21, 2026, citing VulnCheck on 24-hour exploitation)