cd /news/ai-safety/github-and-microsoft-9791-ai-finding… · home › topics › ai-safety › article
[ARTICLE · art-171909] src=explainx.ai ↗ pub= topic=ai-safety verified=true sentiment=↑ positive

GitHub and Microsoft: 9,791 AI Findings Became 99 Verified OSS Reports

GitHub and Microsoft reported on November 17, 2026 that their Codename MDASH pilot, which uses the Microsoft MAI Cyber model, turned 9,791 raw AI vulnerability flags into 99 privately delivered reports across six open-source projects, with 55 fixes completed as of November 15 and 14 already shipped in a release. The pipeline pairs MDASH scanning and fix suggestions with a GitHub Security Lab triage agent built on tailored Copilot skills, then hand verification by seven Microsoft FORGE Lab researchers at roughly two days per project, yielding about one maintainer-facing report per 100 raw flags. GitHub framed the effort as a response to maintainers facing duplicate, low-context AI-generated bug reports, citing Log4j maintainer Christian Grobmeier on the need for defenses to outpace offensive AI.

recommendation

Gate AI-generated security findings behind an automated triage agent for deduplication and validity checks plus human verification before filing private reports; the pilot reduced 9,791 raw flags to 99 reports.

more in playbook: Designing agent systems
by read8 min views2 publishedNov 18, 2026
GitHub and Microsoft: 9,791 AI Findings Became 99 Verified OSS Reports
Image: Explainx (auto-discovered)

On November 17, 2026, GitHub and Microsoft published the results of a Codename MDASH pilot that asks a different question than most AI security marketing: not how many bugs a model can find, but how many verified fixes reach maintainers who are already drowning in AI-generated noise. The official GitHub blog post says thousands of possible issues were flagged; after AI triage and hand review, six projects received 99 private reports with proof-of-concept and suggested patches. As of November 15, maintainers had fixed 55, and 14 fixes were already shipped in a release. Same-day trade coverage—including SecurityBrief's recap—reported 9,791 raw flags citing the pilot figures.

If you run **agents on dependencies** or ship **MCP servers**, this pilot is the counter-pattern to dumping **Scanner-of-the-week PDFs** into public issues. It is also the production sibling to **[MAI Cyber 2 Flash](https://explainx.ai/blog/microsoft-ignite-mai-code-2-seven-models-november-2026)**, which Microsoft announced at **Ignite 2026** while this story was going live.

## TL;DR: what the numbers mean
Question Answer (per GitHub, Nov 17, 2026)
What scanned the code? Codename MDASH usingMicrosoft MAI Cyber model
Raw AI flags Thousands per GitHub; trade press cited9,791
Reports maintainers saw 99 private reports after triage and human verify
Fixed by Nov 15 55 , with14 already in a release
Human cost Seven FORGE Lab researchers,~two days per project
Ratio Roughly one maintainer-facing report per 100 raw flags

Agents that act on your behalf need rules. See how explainx helps secure your agents →

Why maintainers said no to more AI bug reports #

The post opens with a blunt premise: AI made finding vulnerabilities fast and cheap, but fixing still needs human judgment, and in open source that work lands on volunteers. GitHub cites Christian Grobmeier, a Log4j maintainer, saying defenses must beat offensive AI — a line that echoes our stop blocking defenders letter coverage on whether cyber-capable models should reach defenders at all.

Many maintainers already receive duplicates, non-vulnerabilities, and low-context mail. Another raw scan increases triage load without shrinking patch queues. That is why Google d OSS VRP intake when AI-slop reports spiked, and why Warner-style patch rush incidents happen when executives treat every model output as CVE-critical.

The pilot goal was narrower: send only findings that matter, and never send an unverified finding.

The five-step pipeline #

GitHub documents a fixed sequence. Each step has a named owner.

  1. Find.**Codename MDASH scans withMAI Cyber** andsuggests fixes for what it finds.
  2. Filter. AGitHub Security Lab triage agent, built ontailored Copilot skills , removesduplicates , checksvalidity , and focuses onhigh-impact bugs so maintainers never seeraw scan output .
  3. Verify.**Microsoft FORGE Lab researcherscheck every remaining finding by hand** .
  4. Report privately. Reports go through**GitHub private vulnerability reporting** withproof of concept and asuggested fix .
  5. Fix. Maintainers choosewhat ,how , andwhen to merge; work stays onGitHub whereCI ,forks , andadvisories already live.

This is intentionally not a public issue flood. It mirrors how serious vendors want coordinated disclosure, and it aligns with MCP server security guidance that agent tools should fail closed when untrusted input reaches privileged actions.

Maintainer security workflow after AI scan triage for Codename MDASH

Which projects opted in #

All six projects are GitHub Secure Open Source Fund alumni — each historically received $10,000 via GitHub Sponsors plus security training. Maintainers agreed before the scan, reviewed reports privately, and shared qualitative feedback with GitHub:

Project Role in stack Reports mentioned in blog
AutoGPT Agent platform Broken new check caught three days after ship
Caddy Web server 22 reports;all fixed within a week
Express Node.js web framework Reports cut through noise
FastMCP MCP framework Insights beyond prior AI scanners
ImageMagick Image processing 22 reports;two-person team
Netty Java networking Quote: fixing, not finding, is the bottleneck

FastMCP matters for this audience because MCP hosts are now default in Claude Code and Copilot Studio workshops. A verified MCP framework fix is upstream of thousands of agent demos.

What maintainers actually fixed #

GitHub highlights time-to-fix, not model bragging.

Caddy cleared all 22 reports in a week or less, using suggested patches as-is or adapted. AutoGPT had shipped a broken check; the private report landed three days later, pointed at the right line, and the team exceeded the suggested patch, merging to production in under a week. ImageMagick got a report that looked like a familiar class of bug but was not; the summary pointed at exact lines, and maintainers shipped in 15 minutes — versus 30–60 minutes for a typical advisory, per their quote.

Express maintainers said the short list cut noise. Netty stressed that finding is no longer the bottleneck; fixing is. That matches Andy Hall auditor false-positive work on reward hacking and Sapper v2 pentest agents on autonomous testing — discovery tools multiply; merge capacity does not.

Limits GitHub admits upfront #

The blog is explicit that six projects is a small sample. Lessons are early, not proof the method scales everywhere.

Manual verification does not scale yet. Seven researchers times two days per project is incompatible with scanning every popular repo weekly. Meanwhile, low-quality AI reports still hit maintainers daily — the problem the pilot tried to absorb, not abolish.

GitHub also lists open source gaps:

  • Fixes should flow through private vulnerability reporting .
  • Maintainers need CI on private forks to test patches before public merge.
  • Cyber-capable models should verifyproof-of-concept before and after fixes — the same capability debate as**GPT-6 Cyber programs** and**MAI Cyber 1 Flash preview** .

How to protect yourself if you depend on these libraries #

Consumers of AutoGPT, Express, Netty, or ImageMagick should treat November 15 as a checkpoint, not the end of the story.

  1. Turn on Dependabot alerts for repos you ship.
  2. Watch GitHub Security Advisories for the six names;14 fixes were alreadyreleased , but more may land as maintainers finish the remaining44 open reports from the99 batch.
  3. Do not rerun public scanners against those upstreams and file duplicate issues — that recreates the noise this pilot avoided.

For agent builders, use AgentBeam as a runtime guard when agents call shell, git, or package publish tools — separate from static finding.

What people search about Codename MDASH #

Readers landing from Hacker News or GitHub trending often ask whether MAI Cyber is the same product as Codex Security or Defender agents. GitHub’s post draws a bright line: MDASH is a Microsoft Security scanning harness, while GitHub’s triage agent is a Copilot-skills workflow sitting between raw model output and private vulnerability reporting. You do not install MDASH as an npm package; maintainers still merge on GitHub with their usual CODEOWNERS rules.

Another common question is whether 99 reports means 99 CVEs. The blog avoids CVE numbering in the pilot recap; maintainers received private advisories with proof-of-concept and suggested fixes, then chose severity and disclosure timing. 55 fixes by November 15 counts merged remediations, not a public CVE batch count.

Teams comparing Ignite-week announcements should read MAI Cyber 2 Flash preview status before assuming the next open source sweep already runs on Cyber 2.

What GitHub and Microsoft do next #

The post ties forward motion to three threads:

  • Private reporting product — lessons feedGitHub Security Lab investments to reducelow-quality inbound mail.
  • MAI Cyber 2 — Ignite-week model refresh; future scans may use the newer stack described in**Ignite model roundup** .
  • Secure Open Source Fund expansion — future alumni may getchecked AI scans plusCopilot Max , extending the216-project fund narrative (4,000+ vulnerabilities fixed in software downloaded20B+ times monthly, per GitHub).

If you maintain a funded or critical repo, enable gh-secure and private vulnerability reporting now so researchers (human or vendor-run) have a single intake, as Codex Security open source tooling spreads.

What people building AI security should copy #

GitHub ends with a sender standard anyone running agentic scanners should paste into runbooks:

  • Check every finding againstproject knowledge ,threat model , andSECURITY.md .
  • Report privately .
  • Include a suggested fix ; let maintainers set pace.

That is the inverse of leaderboard hacking on ExploitBench-style demos. It also complements external sandboxes for coding agents — isolate execution, but still ship patches through maintainer-owned merge paths.

How this differs from shipping a security model only #

MAI Cyber 2 Flash answers “can a model find bugs?” Codename MDASH answers “can a pipeline deliver mergeable fixes?” That distinction matters for CISOs comparing Microsoft Perception playbooks to AppSec backlog reality.

GitHub’s numbers imply ~1% yield from raw flags to maintainer reports. Security vendors often cite 10x noise reduction; here the denominator is public. If your team runs weekly agent scans, benchmark your signal rate before you promise leadership the same 99-report outcome.

Dependabot and CodeQL still matter: this pilot supplements, not replaces, deterministic rules. Agent autofix features on private repos ride a different rail — maintainer consent plus verified patches.

Supply-chain ripple for MCP and agent frameworks #

FastMCP participation is strategically important. MCP is the default integration spine for Claude Code, Copilot agents, and a growing share of internal tools. A verified fix in FastMCP propagates through thousands of forks faster than a blog CVE.

AutoGPT’s broken-check anecdote shows agent frameworks ship guardrails that themselves become attack surface. Teams running autonomous pentest demos should read Sapper v2 alongside this pilot — offense automation and defense automation share the same maintainers.

Figures and quotes follow GitHub and Microsoft November 17, 2026 blog text unless noted as maintainer quotes.

── more in #ai-safety 4 stories · sorted by recency
── more on @github 3 stories trending now
sponsored brought to you by zahid.host 4,200+ EU-deployed projects
reading about agents? ship yours in a single git push.

Run your AI side-project on zahid.host

EU-based hosting, git-push deploys, automatic HTTPS, no cold starts. Free tier with a custom domain — perfect for shipping the agent you just read about.

$git push zahid main
→ Live at https://your-agent.zahid.host ✓
Get free account → Pricing
from €0/mo · no card required
LIVE [news/github-and-microsoft…] indexed:0 read:8min 2026-11-18 · —