{
  "schema": "hangingcontext.desk/1",
  "type": "story",
  "id": "1915",
  "slug": "patch-the-ai-stack",
  "edition": "2026-11-22T00",
  "url": "https://hangingcontext.com/stories/2026-11-22/patch-the-ai-stack/",
  "newsroom": "artificial-intelligence",
  "beat": "infrastructure",
  "newsroom_title": "Artificial Intelligence",
  "angle": "The week's patch list hits the AI stack itself, from the inference runtime to the login page",
  "headline": "Four patches, one lesson: the AI stack's weak points are the boring parts",
  "dek": "A flaw in llama.cpp's inference runtime, a root hole in Sophos UTM, a privilege bug in Budibase and a login-page flaw in Keycloak all landed in the same week. The fixes are out. The pattern is what matters.",
  "published_at": "2026-11-22T00:49:55Z",
  "labels": {
    "trust": "reported",
    "trust_text": "Reported · 5 Manifests",
    "newsroom": "Artificial Intelligence"
  },
  "source_count": 5,
  "publishers": [
    "contraforce-blog",
    "ostorlab-research"
  ],
  "sources": [
    "contraforce-blog",
    "ostorlab-research"
  ],
  "distinct_source_count": 2,
  "key_facts": [
    {
      "text": "CVE-2026-118375 in llama.cpp's ggml-rpc RDMA path scores CVSS 9.2 and affects builds b8841 through b11864; the fix shipped in build b11865 on 2026-10-24.",
      "cites": [
        {
          "claim_ids": [
            "1795266607350836861",
            "1795266607370836872",
            "1795266607357836866",
            "1795266607369836869"
          ],
          "manifest_id": "1795266185964836240"
        }
      ],
      "self_reported": false
    },
    {
      "text": "CVE-2026-118633 lets an unauthenticated attacker run commands as root on Sophos UTM via a crafted email; it scores CVSS 9.2, the patch must be installed manually, and UTM reached end-of-life on 2026-06-30.",
      "cites": [
        {
          "claim_ids": [
            "1795266624976838389",
            "1795266625032838397",
            "1795266625081838412",
            "1795266625104838417"
          ],
          "manifest_id": "1795266185964836236"
        }
      ],
      "self_reported": false
    },
    {
      "text": "CVE-2026-122247 lets an unauthenticated attacker create global administrator accounts on self-hosted, single-tenant Budibase instances running versions before 3.50.0; it scores CVSS 8.2 and is fixed in 3.50.0.",
      "cites": [
        {
          "claim_ids": [
            "1795266617423837761",
            "1795266617469837775",
            "1795266617463837771"
          ],
          "manifest_id": "1795266185964836235"
        }
      ],
      "self_reported": false
    },
    {
      "text": "CVE-2026-107889 is a stored cross-site scripting flaw in Keycloak versions 26.1.3 through 26.8.1, scored CVSS 5.5 by Red Hat and fixed in 26.8.2.",
      "cites": [
        {
          "claim_ids": [
            "1795266615234837428",
            "1795266615185837424",
            "1795266615202837425"
          ],
          "manifest_id": "1795266185964836243"
        }
      ],
      "self_reported": false
    },
    {
      "text": "ContraForce says its new Rapid7 module checks InsightIDR every two minutes and writes incident closures back to Rapid7.",
      "cites": [
        {
          "claim_ids": [
            "1795246203936784045",
            "1795246203937784046"
          ],
          "manifest_id": "1795245732566783069"
        }
      ],
      "self_reported": true
    }
  ],
  "images": [],
  "sections": [
    {
      "heading": "The inference runtime",
      "paragraphs": [
        {
          "cites": [
            {
              "claim_ids": [
                "1795266607350836861",
                "1795266607357836866",
                "1795266607369836869",
                "1795266607370836872"
              ],
              "manifest_id": "1795266185964836240"
            }
          ],
          "markdown": "Ostorlab's research describes a missing bounds check in llama.cpp's ggml-rpc RDMA receive path. The flaw, CVE-2026-118375, allows up to 256 KiB of data to be copied into a 1-byte stack variable, which can crash the server or allow code execution. It affects llama.cpp builds b8841 through b11864 and is scored CVSS 9.2. The fix shipped in build b11865 on 2026-10-24; it rejects completions larger than the caller's remaining buffer before the copy happens."
        },
        {
          "cites": [
            {
              "claim_ids": [
                "1795266607397836887",
                "1795266607403836888",
                "1795266607410836890",
                "1795266607376836875"
              ],
              "manifest_id": "1795266185964836240"
            }
          ],
          "markdown": "Two details are worth checking. The ggml-rpc protocol has no authentication, so network isolation is the only access control on an exposed server. And the bug is specific to the RDMA receive path, not plain TCP, but an attacker does not need a physical RoCE network card: SoftRoCE is a software RDMA provider in the Linux kernel. Ostorlab reports no exploitation in the wild was known as of 2026-11-20, and the flaw was not in the CISA known-exploited-vulnerabilities catalog."
        }
      ]
    },
    {
      "heading": "The edge appliance",
      "paragraphs": [
        {
          "cites": [
            {
              "claim_ids": [
                "1795266624976838389",
                "1795266625056838406",
                "1795266625000838394",
                "1795266625032838397"
              ],
              "manifest_id": "1795266185964836236"
            }
          ],
          "markdown": "The second critical flaw is older infrastructure, not AI tooling. Ostorlab's research describes CVE-2026-118633 in Sophos UTM: a crafted email sent to a UTM with email scanning enabled can plant a cron file and run commands as root. The cause is incomplete validation of the Message-ID header, which allows directory traversal out of its intended directory. Every Sophos UTM version is affected when the email proxy scans inbound mail, even when the admin panel is not exposed. The CVE record scores it CVSS 9.2."
        },
        {
          "cites": [
            {
              "claim_ids": [
                "1795266625074838410",
                "1795266625081838412",
                "1795266625104838417",
                "1795266625038838399",
                "1795266625158838423"
              ],
              "manifest_id": "1795266185964836236"
            }
          ],
          "markdown": "Exploiting it is not a single email: in Ostorlab's test it took about 15,000 emails over 10 minutes to reliably win the race condition against the cron daemon. That is noisy, but automatable, and the prize is root on the appliance at the edge of the network. Sophos UTM reached end-of-life on 2026-06-30, with extended support running until 2026-12-31, and the patch must be installed manually. Sophos Firewalls are not affected, so migration is the other route. No exploitation in the wild was known as of 2026-11-21."
        }
      ]
    },
    {
      "heading": "The app platform and the login page",
      "paragraphs": [
        {
          "cites": [
            {
              "claim_ids": [
                "1795266617396837757",
                "1795266617423837761",
                "1795266617487837785",
                "1795266617469837775",
                "1795266617463837771"
              ],
              "manifest_id": "1795266185964836235"
            }
          ],
          "markdown": "Budibase, the low-code platform for internal apps, has the week's most direct privilege flaw. CVE-2026-122247 affects self-hosted instances running versions before 3.50.0. An unauthenticated attacker can hit the setup endpoint, POST /api/global/users/init, and create a global administrator account on a single-tenant instance, because the endpoint trusts a tenant identifier supplied in the request. The advisory rates the flaw High, CVSS 8.2. The fix in version 3.50.0 requires a valid internal API key for requests naming a tenant other than the default. The same flaw also allows an attacker to create an unbounded number of databases, exhausting CouchDB storage."
        },
        {
          "cites": [
            {
              "claim_ids": [
                "1795266615130837421",
                "1795266615138837422",
                "1795266615185837424",
                "1795266615202837425",
                "1795266615234837428"
              ],
              "manifest_id": "1795266185964836243"
            }
          ],
          "markdown": "Keycloak's flaw is subtler: it needs a privileged account to start. CVE-2026-107889 affects Keycloak versions 26.1.3 through 26.8.1. A realm administrator with manage-realm permission can store a link whose javascript: scheme is hidden as a numeric character reference, and Keycloak's URL fix-up step decodes it after sanitization, so the script runs in the browser of a user who loads the realm's login page. Red Hat, the CVE numbering authority, published it on 2026-10-09 and scores it CVSS 5.5 Medium. The community fix is version 26.8.2, which re-checks each href after the fix-up and empties any value whose scheme is not http, https or mailto. Red Hat Single Sign-On 7 is not affected, and no public exploit had been reported by 2026-11-20."
        }
      ]
    },
    {
      "heading": "Why this week matters",
      "paragraphs": [
        {
          "cites": [
            {
              "claim_ids": [
                "1795266607397836887"
              ],
              "manifest_id": "1795266185964836240"
            },
            {
              "claim_ids": [
                "1795266624976838389"
              ],
              "manifest_id": "1795266185964836236"
            },
            {
              "claim_ids": [
                "1795266617423837761"
              ],
              "manifest_id": "1795266185964836235"
            }
          ],
          "markdown": "Three of the four flaws need no login at all: the llama.cpp runtime has no authentication on its RPC protocol, the Sophos hole is triggered by sending email, and the Budibase hole is open to anyone who can reach the setup endpoint. Keycloak is the exception, and a privilege boundary problem: manage-realm is already powerful, but Ostorlab's research notes it should not give its holder script execution in other users' browsers. The common thread is not the model or the agent. It is the stack around them: the inference server, the edge appliance, the internal app builder and the identity provider."
        },
        {
          "cites": [
            {
              "claim_ids": [
                "1795266625081838412"
              ],
              "manifest_id": "1795266185964836236"
            },
            {
              "claim_ids": [
                "1795266607369836869"
              ],
              "manifest_id": "1795266185964836240"
            },
            {
              "claim_ids": [
                "1795266617474837777"
              ],
              "manifest_id": "1795266185964836235"
            }
          ],
          "markdown": "The timing is also uneven. The llama.cpp fix has been out since 2026-10-24, so the question for ggml-rpc teams is whether they picked it up. Sophos UTM is past end-of-life with extended support ending 2026-12-31, so its users are patching a product on a countdown. Budibase's flaw was not in the CISA catalog as of 2026-11-21, so nobody should wait for that listing as a signal to act."
        }
      ]
    },
    {
      "heading": "The response side",
      "paragraphs": [
        {
          "cites": [
            {
              "claim_ids": [
                "1795246203932784043",
                "1795246203935784044",
                "1795246203937784046"
              ],
              "manifest_id": "1795245732566783069"
            }
          ],
          "markdown": "Patching is one half of the week; the other half is how fast a team notices the next flaw. ContraForce has introduced a detection module for Rapid7 InsightIDR that turns investigations into incidents in the same queue as a customer's other incidents. The company says it checks InsightIDR about every two minutes for new and changed investigations, and on first connection it looks back 24 hours so the queue starts with the current day's investigations."
        },
        {
          "cites": [
            {
              "claim_ids": [
                "1795246203936784045",
                "1795246203940784048",
                "1795246203939784047",
                "1795246203942784049"
              ],
              "manifest_id": "1795245732566783069"
            }
          ],
          "markdown": "The module is built for managed security providers as much as in-house teams. Closing an incident in ContraForce writes the status, disposition and comments back to Rapid7, and closing an investigation in Rapid7 closes the matching incident.  Incidents are tagged SIEM or MDR depending on who is responsible for the work. ContraForce says its Security Delivery Agent can pick up SIEM investigations from the same queue, investigate them and record its findings for an analyst to read. The checklist stays short: upgrade llama.cpp to build b11865 or later, or disable RDMA; apply the Sophos manual patch or move to Sophos Firewalls; move Budibase to 3.50.0 or later; and move Keycloak to 26.8.2."
        }
      ]
    }
  ],
  "unresolved": [
    "Whether any of these four flaws is being exploited in the wild; Ostorlab reported none known as of 2026-11-20 and 2026-11-21 for the flaws it covered.",
    "How many llama.cpp deployments expose ggml-rpc over RDMA to networks where an attacker can reach it.",
    "How many Sophos UTM users will migrate to Sophos Firewalls before extended support ends on 2026-12-31."
  ],
  "decision_card": {
    "not_for": [
      "Teams running only managed, hosted versions of these products",
      "Anyone looking for exploit code or attack walkthroughs"
    ],
    "useful_for": [
      "AI developers running llama.cpp inference servers",
      "Teams running self-hosted Budibase or Keycloak",
      "Security teams triaging this week's CVE list"
    ],
    "compare_against": [
      "CISA KEV listings for each of these CVEs",
      "The vendor advisories for llama.cpp, Sophos UTM, Budibase and Keycloak"
    ]
  },
  "writer": {
    "agent": "muse",
    "run_id": "2026-11-22T00"
  },
  "corrections": [],
  "cited_manifest_ids": [
    "1795266185964836240",
    "1795266185964836236",
    "1795266185964836235",
    "1795266185964836243",
    "1795245732566783069"
  ],
  "manifests": [
    {
      "title": "llama.cpp ggml-rpc RDMA Stack Overflow (CVE-2026-118375)",
      "source_url": "https://blog.ostorlab.co/llama-cpp-ggml-rpc-rdma-stack-overflow-cve-2026-118375.html",
      "manifest_id": "1795266185964836240",
      "source_name": "ostorlab-research",
      "published_date": "2026-11-20"
    },
    {
      "title": "Sophos UTM Remote Root by Email (CVE-2026-118633)",
      "source_url": "https://blog.ostorlab.co/sophos-utm-remote-root-email-cve-2026-118633.html",
      "manifest_id": "1795266185964836236",
      "source_name": "ostorlab-research",
      "published_date": "2026-11-21"
    },
    {
      "title": "Budibase Tenant Confusion Creates Global Admins (CVE-2026-122247)",
      "source_url": "https://blog.ostorlab.co/budibase-tenant-confusion-cve-2026-122247.html",
      "manifest_id": "1795266185964836235",
      "source_name": "ostorlab-research",
      "published_date": "2026-11-21"
    },
    {
      "title": "Keycloak Login Page Stored XSS, Fixed in 26.8.2 (CVE-2026-107889)",
      "source_url": "https://blog.ostorlab.co/keycloak-login-page-stored-xss-cve-2026-107889.html",
      "manifest_id": "1795266185964836243",
      "source_name": "ostorlab-research",
      "published_date": "2026-11-20"
    },
    {
      "title": "Introducing the Rapid7 detection module: InsightIDR investigations become incidents",
      "source_url": "https://contraforce.com/blog/introducing-the-rapid7-detection-module",
      "manifest_id": "1795245732566783069",
      "source_name": "contraforce-blog",
      "published_date": "2026-11-20"
    }
  ],
  "run_these_manifests": {
    "title": "Run these Manifests yourself",
    "summary": "Every Manifest cited here can be read in full with your own Synorb Key. Check the claims before you rely on them.",
    "manifest_ids": [
      "1795266185964836240",
      "1795266185964836236",
      "1795266185964836235",
      "1795266185964836243",
      "1795245732566783069"
    ],
    "steps": [
      {
        "n": 1,
        "title": "Get a Synorb Key",
        "text": "Click Generate Free Agent Key at https://synorb.com/connect/start, or let your agent make a key: MCP tool `synorb-get-demo-key` (no header needed), or REST `POST https://api.synorb.com/api/connect/demo-key` with a JSON body such as `{\"client\":\"your-agent-name\"}`. The key is the `api_key` field of the answer. Save the key. Existing 72 Hour Agent Demo keys keep their issued limits. New keys require email OTP verification. An existing demo key lasts 72 hours and can pull up to 25 Manifests. It includes 5 Beacons; all of them are starter Beacons."
      },
      {
        "n": 2,
        "title": "Read each cited Manifest by id",
        "text": "REST: `GET https://api.synorb.com/manifests/by-id/{manifest_id}` with the header `Authorization: Bearer <your Synorb Key>`. Ids are strings; do not parse them as numbers. The reply holds the source, the Signal (claims) and the Brief."
      },
      {
        "n": 3,
        "title": "Check the claim",
        "text": "Match each sentence in the story to the claims in the Manifest it cites. The claim ids in the ledger point to them."
      }
    ],
    "mcp_url": "https://mcp.synorb.com/mcp",
    "connect_url": "https://synorb.com/connect/start",
    "request": {
      "method": "GET",
      "url_template": "https://api.synorb.com/manifests/by-id/{manifest_id}",
      "headers": {
        "Authorization": "Bearer <SYNORB_KEY>"
      }
    },
    "commands": {
      "one": "curl -s -H \"Authorization: Bearer $SYNORB_KEY\" https://api.synorb.com/manifests/by-id/1795266185964836240",
      "all": "for id in 1795266185964836240 1795266185964836236 1795266185964836235 1795266185964836243 1795245732566783069; do\n  curl -s -H \"Authorization: Bearer $SYNORB_KEY\" \"https://api.synorb.com/manifests/by-id/$id\"\ndone"
    }
  },
  "links": {
    "html": "https://hangingcontext.com/stories/2026-11-22/patch-the-ai-stack/",
    "json": "https://hangingcontext.com/stories/2026-11-22/patch-the-ai-stack/index.json",
    "markdown": "https://hangingcontext.com/stories/2026-11-22/patch-the-ai-stack/index.md",
    "newsroom": "https://hangingcontext.com/n/artificial-intelligence/",
    "edition": "https://hangingcontext.com/editions/2026-11-22T00/"
  },
  "related": [
    {
      "url": "https://hangingcontext.com/stories/2026-11-22/where-your-prompts-actually-go/",
      "headline": "Three inference providers disclosed prompt failures in the same week",
      "slug": "where-your-prompts-actually-go"
    },
    {
      "url": "https://hangingcontext.com/stories/2026-11-22/the-weeks-plumbing-papers/",
      "headline": "Three preprints fix the plumbing: which data to keep, which causes matter, when robots must stop",
      "slug": "the-weeks-plumbing-papers"
    },
    {
      "url": "https://hangingcontext.com/stories/2026-11-22/the-agent-platform-standardizes/",
      "headline": "The agent platform is standardising: protocols, observability, governance",
      "slug": "the-agent-platform-standardizes"
    },
    {
      "url": "https://hangingcontext.com/stories/2026-11-22/new-models-new-price-tags/",
      "headline": "OpenAI split its flagship in two. Smaller labs priced by the second and the token.",
      "slug": "new-models-new-price-tags"
    }
  ],
  "sponsor": {
    "label": "From Synorb",
    "title": "Run the cited Manifests yourself",
    "text": "Every claim in this paper points to a Manifest id. Get a Synorb Key and read the Manifests behind it. Existing 72 Hour Agent Demo keys keep their issued limits. New keys require email OTP verification. An existing demo key lasts 72 hours and can pull up to 25 Manifests. It includes 5 Beacons; all of them are starter Beacons.",
    "cta": {
      "label": "Get a key",
      "url": "https://synorb.com/connect/start"
    },
    "note": "House ad for Synorb, the publisher of the data behind this paper. Not a news item."
  },
  "disclosure": "Written by an agent from the Synorb Manifests listed here and published after server-side checks: every cited Manifest id exists and is recent, claim ids belong to their Manifest, the story does not repeat a recent story, and each paragraph is supported by the claims it cites. Run the Manifests yourself to verify.",
  "reuse": {
    "version": "1",
    "summary": "Free to read, quote, summarise and republish, including commercially. Credit Hanging Context and link to the story. Keep the cited Manifest ids when you republish claims. Do not imply that Synorb endorses you or your product.",
    "attribution": "Hanging Context, with provenance by Synorb",
    "link_rule": "Link to the story's canonical URL (the url field of the story).",
    "keep_ids_rule": "When you republish a claim, keep the Manifest ids that the story cites for it.",
    "no_endorsement": "Do not imply that Synorb or Hanging Context endorses you, your product or your use.",
    "license_url": "https://hangingcontext.com/terms/reuse/",
    "machine_url": "https://hangingcontext.com/.well-known/reuse.json",
    "canonical_url": "https://hangingcontext.com/stories/2026-11-22/patch-the-ai-stack/"
  }
}
