# AI Exploit Reports: A Patch-Triage Workflow That Starts with Your Deployed Version

Separate a credible vulnerability, an affected system and an unverified exploit claim before deciding what to patch first.

By Paras Madan | 2026-11-17

Google's November 16 AI Threat Tracker describes attackers automating patch analysis and exploitation work. It also warns that some LLM-generated exploits claimed to work do not. For a team running public applications, the useful response is to verify affected deployments quickly, not wait for a convincing exploit video.

The [GTIG report](https://cloud.google.com/blog/topics/threat-intelligence/agentic-threat-activity-crosses-a-significant-threshold) describes an exposed staging server hosting a framework called `exploit_agent`, configured to compare software changes, identify potential vulnerabilities and execute generated exploits. Google also found victim data and related tooling. That evidence does not provide a success rate for every generated exploit, or make every social-media reproduction trustworthy.

Here is a documentation-based triage workflow I would use. The example below is fictional, and I have not tested a vulnerability or patched a live service for this article.

## Build one record per affected deployment

Start with the vendor advisory and your asset inventory. The [CERT/CC SSVC deployer model](https://certcc.github.io/SSVC/howto/deployer_tree/) explicitly connects deployed product versions with supplier-provided remediations or mitigations. Its unit of work is a remedy applied to a particular product instance, rather than a vulnerability identifier floating in a spreadsheet.

For an initial ticket, collect these fields:

| Field | What to record |
| --- | --- |
| Advisory | Original vendor URL and revision date |
| Deployment | Service name, owner and environment |
| Installed state | Actual running version or build identifier, with evidence |
| Applicability | Affected version range and required configuration |
| Exposure | Who can reach the affected feature and how |
| Response | Vendor fix or mitigation, owner and review time |
| Verification | How you will confirm the change reached the running service |

If the version is unknown, leave it unknown and assign someone to resolve it. Check the deployed artifact or supported version-reporting mechanism; a repository dependency is not necessarily the version serving traffic.

## Work through a concrete example

Imagine an advisory for a fictional document portal, PaperDesk. It says versions 4.2.0 through 4.2.3 are affected when an optional preview feature is enabled, and identifies 4.2.4 as the fix.

Your repository already declares 4.2.4, but the production deployment record shows 4.2.2. The preview feature is enabled and reachable by external customers. That is enough to open an affected-deployment ticket. A failed reproduction copied from a forum would not override the version and configuration evidence.

Now consider a second instance used for internal testing. It runs 4.2.2 but has the feature disabled. Record that difference and verify the advisory's condition instead of marking both instances identically. Do not infer that “internal” means unreachable or harmless; document the actual access path.

## Separate priority from the loudness of the claim

SSVC considers exploitation evidence, system exposure and impact when prioritizing a deployer's response. A public exploit claim and observed exploitation are different inputs. Preserve that distinction in the ticket instead of treating repost count as confirmation.

I would give the incident owner two separate questions: “Are we affected?” and “What evidence changes the urgency?” A vendor confirmation, a trusted incident report or relevant activity in your own logs can change the second answer. The first still needs the installed version and configuration.

Do not run an unknown exploit against production to settle the argument. Use approved validation methods and the vendor's guidance. If compromise is suspected, involve incident response; applying a patch alone does not establish that unauthorized access has ended.

## Project Zero's disclosure clock is not your patch window

Project Zero's [November 13 policy change](https://projectzero.google/2026/11/disclosure-policy.html) replaces its previous 90-day remediation period plus 30-day rollout allowance with disclosure 90 days after notification. It plans to shorten that interval to 60 days on February 2, 2027, and 30 days on January 3, 2028. Exceptions remain, including active exploitation and already-public details.

Those are Project Zero's disclosure rules, not a universal deadline or permission to leave an affected service unpatched for 90 days. The policy separately emphasizes the gap between a fix becoming available and users receiving it. Your deployment record should track that second clock.

## Close on evidence, not a successful pipeline

Before closing the ticket, confirm the fixed build is running on each affected instance and check the relevant functionality after rollout. Keep any temporary mitigation and its removal conditions in the same record. 

The practical advantage is shorter uncertainty: knowing which service needs action, who owns it and what proves the action worked. That remains useful whether the next exploit was written by a person or generated by an agent.

I'm Paras. I work at the intersection of AI, engineering and distribution.

I help technical teams turn security findings and product changes into clear, practical developer content.

[Book a 30-minute call](https://cal.com/paras-madan/30min)
