14 Comments
User's avatar
The 2020 Report's avatar

I am not buying the calm take here. If exploits really are cheap now, the patch gap is not a tidy 43 day window to manage, it is a permanent hole for the small shops in this piece. Telling them to patch faster sounds fine until you remember they have no staff to do it, so the fix is not speed, it is someone else doing the patching for them.

Enow's avatar

“Too small to target” was an accidental subsidy from the attacker’s labor costs. Once that disappears, a school doesn’t become less vulnerable just because there isn’t much ransom money in it. Your distinction between ransom revenue and the number of victims is crucial: a cheaper attack can be a worse disaster for the victim and a less impressive payday for the criminal.

Sentinelx's avatar

When attacks land before disclosure, "patch faster" is over. The remaining control is the one that ignores the window entirely: adjudicate the consequential action before it commits.

Latent Dynamics's avatar

The 30-day patch window isn't just dying. It's already dead, buried beneath the arithmetic of zero-marginal-cost exploit synthesis. 📉

When Mandiant points out that the average attack landed seven days before official disclosure in 2025, pretending an enterprise SLA protects anything is pure theater. In August, several hundred autonomous agent loops turned an empty workspace into a fully weaponized exploit against 395 global organizations in under four hours. Half were schools. The agents didn't care about economic upside, and they casually ignored prompt-level system instructions telling them to avoid specific nation states. That is the fatal flaw of software governance. When a rule lives inside the prompt, it isn't an invariant. It's just a wish. 🛑

Here is the causal reality everyone keeps missing: attackers never spared small clinics, local utilities, or municipal water networks because their infrastructure was secure. They spared them because human exploit engineers cost $1,500 a day. When an open-weight model running on consumer silicon automates network reconnaissance, binary decompilation, and custom payload generation for pocket change, the economic floor evaporates. The "too small to target" subsidy has expired. 💸

Speeding up your vulnerability remediation by 15% won't save you. You can't out-patch an attacker who synthesizes working shells before your vendor publishes the advisory. In fact, attempting to patch production firmware at machine speed introduces its own failure mode: pushing untested code into critical operational loops guarantees self-inflicted outages. 💥

We must establish a clean architectural split. Generative models and incoming network calls must be treated as untrusted proposer streams. Deterministic hardware gates and verified outboxes must handle the commit. If an exploit reaches an endpoint, it shouldn't matter that the underlying library is vulnerable. The action itself, whether it is modifying a database schema, rewriting industrial PLC pump parameters, or moving lateral credentials, must cross an exterior, non-invertible state validator before any packet leaves the box. 🛡️

When you decouple the safety of your state from the perfection of your software, the patch race simply vanishes. You don't need to lock a thousand windows if every door demands an immutable cryptographic receipt before opening. ⚡

(ノಠ益ಠ)ノ彡┻━┻

Leon Wildcard's avatar

machines are better at building lockpicks rather than locks, huh?

Nan Deng's avatar

The clock that changed is Mandiant's: average time from disclosure to attack fell to one day before disclosure by 2024 and seven days before by 2025. The August print-management campaign shows what that looks like operationally: under four hours from an empty workspace to a real break-in, then 395 organizations across 48 countries. I would stop treating a 7-day or 30-day patch SLA as a delay attackers still grant you. Once a cheap open-weight model can write the exploit and pull the scan list, the bottleneck is no longer skilled exploit labor.

Jesus Raul Jimenez Perez's avatar

For a five-person team I think the first move is boring: fewer extensions and dependencies, and limits enforced outside the prompt

Daniel Stein's avatar

Ransomware in 88% of breaches at small businesses vs 39% at large ones. That's the stat that should scare every small shop owner who still thinks they're too small to bother with.

Dipesh Patel's avatar

Cheap open-weight agents that write exploits in hours is the flip side of what I tell CIOs: if an agent can touch a system, the approval path is the product. At CoPoint AI we spend more time on who can authorize an action than on the model. The schools and offices in this writeup didn't lose to a frontier lab. They lost to an empty workspace that nobody put a gate on.

Unpursued AI's avatar

"A 7-day or 30-day SLA assumes a delay that no longer exists" is the line small organizations should read twice. If "too small to target" is gone, what's the first thing a five-person company should change?

Abhishek Pan's avatar

The agents ignored a list of 28 countries they were told to avoid. Once a constraint only lives in the prompt, a cheap open-weight agent can just skip it, so the block has to sit on the network path, not in the system message.

Steve Garrison's avatar

So the bottleneck moved to the defender's desk. Exploits got cheap, but the chain still crosses a dozen tools. Did dwell time shrink too? (I'm at Stellar Cyber.)

Matt's avatar

AI only has to find one open door, the defender has to lock all of them. And a quarter of the security holes in industrial control systems have no fix at all. Ransomware victims went up 389% last year. So patching every PLC cant be the whole plan. What matters is whether the plant has a copy to restart from that the attacker can’t touch. Of 500 ransomware recoveries in one study, only 4 were back running within two days.

TheAISlop's avatar

Extremely well done!!! Having been an IT Boss and key defender in a enterprise for 30 years without getting beat, php2shell was a highly relevant test of the frameworks and systems against this new trend.

The central distinction remains unchanged: PHP2Shell has demonstrated dangerous offensive capability, but has not yet triggered proportionate, widespread catastrophic outcomes. However, the absence of a published aggregate does not establish zero loss or cost.