Most teams can tell you their deploy frequency, their change failure rate and roughly how long a pull request waits for review. Far fewer can tell you how many days it took to get the last critical vendor security update into production.
That number matters more this month than it did last month.
This post covers what Google's threat intelligence team reported, what it means for engineering teams, and a log template you can start filling in today.
What was reported
Google Threat Intelligence Group, "GTIG AI Threat Tracker: Agentic Threat Activity Crosses a Significant Threshold", 16 Nov 2026.
- GTIG says that in the past quarter it saw attackers use AI frameworks to "fully automate vulnerability discovery and exploitation without a human in the loop for the first time".
- The example, a framework called
exploit_agent, runs continuous differential analysis on popular web software. It compares code before and after a patch, works out the flaw, drafts an exploit and executes it. - One figure shows it inspecting access checks in a WordPress plugin. Plugins count.
- In an August case, researchers and Huntress used LLMs to build working exploits for two PaperCut flaws within a day of disclosure. GreyNoise reported a campaign against hundreds of organisations four days after disclosure (as summarised by GTIG).
- Separately, an intrusion set modified
.claude/settings.jsonin internal repos so Claude Code downloaded a backdoor on start-up.
Mandiant time-to-exploit (TTE), defined as the average time to exploit a vulnerability before or after its patch is released:
| Period | Average TTE | Source |
|---|---|---|
| 2018 to 2019 | 63 days | Mandiant, Oct 2024 |
| 2021 to 2022 | 32 days | Mandiant, Oct 2024 |
| 2023 | 5 days (47 with 15 outliers kept) | Mandiant, Oct 2024 |
| 2025 investigations | about minus 7 days (estimate) | M-Trends 2026 |
The samples differ by report, so read these as snapshots rather than one trend line. Means are also easy to skew. Mandiant's 2023 split is more useful: 12% of n-days were exploited within one day of the patch, 29% within a week and 56% within a month.
What this means for engineering teams
Everything in this section is my inference, not something the report prescribes.
A security patch is now a disclosure. Once the fixed version ships, the diff tells anyone what the old version got wrong. Automated tooling reads diffs very fast, so your exposure clock starts at vendor release, not when your scanner raises a ticket.
Your patch latency is a production metric. Treat "days from vendor release to running in prod" for internet-facing components like any other SLO. Track the median and the worst case. The worst case is what gets you breached.
Normal release cadence is the wrong vehicle. If a critical fix for an exposed service waits for the next sprint, it inherits the sprint's timeline. Pre-agree a fast path: criteria, approver, minimum smoke tests, rollback. Cyber Essentials allows 14 days for critical and high fixes. Read that as the outer limit.
Dependency bots help less than you'd hope. Dependabot or Renovate will open the PR. They won't get it through a busy review queue, a booked staging environment and a change window. Measure the whole path.
Agent config files are code. If a settings file can make your coding assistant run a command on start-up, it belongs under CODEOWNERS and review like a CI workflow.
The artefact: a patch latency log
Keep this in your repo or ops wiki. One row per critical or high fix on anything internet-facing. The example rows are made up to show the format.
## Patch latency log (internet-facing only)
| Component | Advisory / CVE | Vendor release | Prod deployed | Days | Waited on | Mitigation before patch | Compromise check |
|---|---|---|---|---|---|---|---|
| EXAMPLE: reverse proxy | Vendor advisory, CVSS 9.1 | 2026-11-03 | 2026-11-04 | 1 | Nothing, fast path | None needed | Logs reviewed, clean |
| EXAMPLE: CMS plugin | Plugin changelog, "security fix" | 2026-11-05 | 2026-11-19 | 14 | Staging booked for UAT | WAF rule on admin path | Not done (action) |
| EXAMPLE: VPN appliance | Vendor PSIRT, exploited | 2026-11-10 | 2026-11-12 | 2 | Change board | Disabled web portal | IOC sweep, clean |
**This month:** median __ days, worst case __ days, rows over 14 days: __
### Fast path criteria (agree before you need them)
- [ ] Internet-facing AND vendor says critical/high, or CVSS v3 >= 7, or no severity given
- [ ] Named approver on call: ________
- [ ] Minimum tests: ________
- [ ] Rollback plan tested in the last 90 days
- [ ] Mitigation options listed per component (feature flag, WAF rule, take offline)
Two fields do most of the work. "Waited on" shows you where the days go, which is rarely the patch itself. "Compromise check" exists because a negative average TTE means some exploitation happens before any fix exists, so patching closes the door but doesn't tell you whether someone already walked through it.
Caveats
- GTIG's report is one vendor's telemetry. Its exploit_agent example is a single framework found on one server.
- The PaperCut timings come via GTIG's summary of Huntress and GreyNoise, not their original write-ups.
- Mandiant's TTE figures use different samples, and the 2025 figure is an estimate from incident investigations.
- Faster patching has real costs. A bad update can cause its own outage, which is why the fast path needs tests and a rollback as well as speed.
Over to you
What is the worst-case patch latency for an internet-facing component in your stack right now, and what did that fix spend most of its time waiting on?
Top comments (0)