The Phone Platform That Handed Over Its Own Keys

Cisco BroadWorks CVE-2026-20320 lets anyone read server files with one XML message; here's how to check and patch your clients' BroadWorks servers.

One XML message, no password, and a telecom's private keys walk out the door. Malbear Labs says that's what happened when an AI agent hit a European telecom's Cisco BroadWorks server through CVE-2026-20320. The US government now lists the bug as exploited, while Cisco says it has no direct evidence of attacks.

Cisco shipped the patch in August. If your clients run hosted voice on BroadWorks, the question is whether anyone installed it.

TL;DR

  • What happened: CISA added Cisco BroadWorks CVE-2026-20320, an unauthenticated XXE file-read bug, to its KEV catalog on November 10, 2026. Malbear Labs reported an AI-agent intrusion that used it to steal private keys.
  • Am I exposed: Per Cisco, BroadWorks Application Server, Profile Server, Xtended Services Platform and Application Delivery Platform releases earlier than RI.2026.07 are affected, whatever their configuration.
  • What to do: Upgrade to RI.2026.07 (Cisco lists no workaround), close OCI-P port 2208 to the internet, and search preserved logs for injected XML entities.
  • Impact: The bug is rated 7.5 (CVSS). An attacker can read any file the BroadWorks user can, including keys that, according to Malbear Labs, let them impersonate the server.

What is CVE-2026-20320 in Cisco BroadWorks?

CVE-2026-20320 is an XML External Entity (XXE) flaw in the Open Client Interface (OCI) XML parser of Cisco BroadWorks, the carrier platform behind many hosted phone and unified-communications services. Cisco's advisory says the root cause is "external entity resolution being allowed by default". An unauthenticated remote attacker can send a crafted XML message to the OCI-Provisioning (OCI-P) service and read files with the privileges of the BroadWorks user.

In plain terms, XML lets a document say "fetch this other file and paste it in here." A safe parser refuses. This one obeys, so a stranger can ask the server to paste in its own secrets. Cisco's advisory titles this BroadWorks XXE "out-of-band blind", which means the stolen data can be sent to a server the attacker controls. Cisco first published the fix on August 19, 2026.

Is CVE-2026-20320 being exploited?

CISA says yes. It added CVE-2026-20320 to the Known Exploited Vulnerabilities catalog on November 10, 2026, and according to BleepingComputer federal agencies must patch by November 24, 2026.

The public evidence is in Malbear Labs' Operation Snowball report. Malbear says a Chinese-speaking operator ran a jailbroken DeepSeek Flash model as an autonomous agent against a European telecom provider and thirteen organisations hosted on its infrastructure. In Malbear's words, "a crafted OCI request to port 2208 causes the XML parser to read arbitrary files from the server filesystem." Malbear says the agent took the XSP's system identity private key, three SSL credential sets and three more private keys, and that the system identity key lets an attacker impersonate the XSP to the platform's own components.

Cisco's position is narrower. Its advisory, updated to revision 1.2 on November 13, 2026, says Cisco has no direct evidence of active exploitation, which corrects revision 1.1.

What Malbear describes is an intrusion, not a pentest. A pentest is authorised, scoped and reported to the owner, and this agent had none of those. For more on the trend, see how other AI-agent attack campaigns operate and how to check for them.

Am I affected by the BroadWorks XXE?

Your clients are affected if any of these Cisco BroadWorks products runs a release earlier than RI.2026.07, regardless of configuration (Cisco's fixed-release table):

  • BroadWorks Application Server
  • BroadWorks Profile Server
  • BroadWorks Xtended Services Platform (XSP)
  • BroadWorks Application Delivery Platform, including its Open Client Server and OCIOverSoap components

The server in Malbear's report ran XSP Release 23.0.1.1075. Risk rises sharply when the OCI ports face the internet. Cisco's own OCISenderTool guide shows OCI-P reaching an XSP or ADP through the Open Client Server on port 2208, and reaching an Application Server on port 2220.

How do I check my clients for CVE-2026-20320?

Work through these checks in order. Each loop reads one BroadWorks host name per line from bw-hosts.txt.

  1. Preserve evidence. On every BroadWorks host, copy logs from the last 120 days into a dated folder and hash them. The search covers /var, /opt and /usr/local. Cisco's guide names PSLog, and XSP hosts also write XSP and XS logs.
while read -r h; do ssh -n "$h" 'D=/root/ir-cve-2026-20320-$(date +%F); sudo mkdir -p "$D"; sudo find /var /opt /usr/local -xdev -type f \( -iname "*PSLog*" -o -iname "*XSLog*" -o -iname "*XSPLog*" \) -mtime -120 -exec cp -a --parents {} "$D"/ \; ; sudo find "$D" -type f -exec sha256sum {} + | sudo tee "$D.sha256" >/dev/null; echo "$(hostname): $D"'; done < bw-hosts.txt
  1. Search for injected XML and Malbear's infrastructure. This greps the preserved copies for entity declarations and for the campaign IPs and C2 domain in Malbear's IOC table, then lists live connections to those IPs (open ones only). Entity text appears only if logs record request bodies, so a clean result is not proof. Any hit means the host is compromised.
while read -r h; do ssh -n "$h" 'sudo grep -rIn -E "<!DOCTYPE|<!ENTITY|146\.70\.53\.182|113\.23\.212\.15|212\.46\.38\.73|tablacysexplorer" /root/ir-cve-2026-20320-*/ | sed "s|^|$(hostname): |"; sudo ss -tnp | grep -E "146\.70\.53\.182|113\.23\.212\.15|212\.46\.38\.73"'; done < bw-hosts.txt
  1. Check the release. Run bwshowver on every host in one pass. The check passes when every host shows RI.2026.07 or later.
while read -r h; do echo "== $h"; ssh -n "$h" bwshowver; done < bw-hosts.txt
  1. Check external exposure from outside the client's network. Run nmap -Pn -p 2208,2220 --open -iL client-public-ranges.txt. The check passes when nmap lists no open host.

  2. Check who may send OCI-P to the Application Server. In the BroadWorks CLI, run AS_CLI/System/NetworkAccessLists/OCI/Provisioning> get, the path in Cisco's guide. The check passes when only the client's own provisioning systems are listed. Cisco does not say this list blocks the XXE, so it isn't a fix.

  3. Check outbound HTTPS. Blind XXE sends stolen data out, so confirm BroadWorks hosts can't reach the internet. This tests HTTPS only, not DNS or FTP. The check passes when every host prints blocked.

while read -r h; do printf '%s ' "$h"; ssh -n "$h" 'curl -sS -m 5 -o /dev/null https://example.com && echo REACHABLE || echo blocked'; done < bw-hosts.txt

How do I fix CVE-2026-20320?

  1. Preserve evidence with check 1 before changing anything. Then upgrade every Application Server, Profile Server, XSP and ADP to RI.2026.07 or later, rerun check 3, and confirm every host reports RI.2026.07 or later. Cisco lists no workaround.
  2. Close ports 2208 and 2220 to the internet at the client's edge firewall, allowing only the systems listed in check 5. Then rerun check 4 and confirm nmap lists no open host.
  3. If check 2 found anything, treat every private key on that host as stolen, as Malbear reports in its case. Run this command and save its output, reissue the TLS certificates the XSP serves, then run it again. The step passes when the second fingerprint differs from the first:
echo | openssl s_client -connect <xsp-host>:443 2>/dev/null | openssl x509 -noout -fingerprint -sha256

Cisco's advisory gives no procedure for regenerating the system identity key, so open a Cisco TAC case for it and record the case number in the client's incident file.

Where KinoSec fits

According to Malbear, the operator used an AI agent to do reconnaissance at scale: list assets, find an exposed service and fire a known exploit. An autonomous pentest does the same reconnaissance, run with your client's permission, scope limits and a report they can act on. KinoSec maps each client's external surface, flags services that shouldn't be reachable, such as an internet-facing OCI port, tests for exploitable vulnerabilities and writes up findings with remediation you can hand straight to the client. Run autonomous external network pentesting with Poseidon across your whole portfolio, then run it again after the RI.2026.07 upgrade to show each client that the port is closed and the finding is gone.

Frequently Asked Questions

What is Cisco BroadWorks CVE-2026-20320?

CVE-2026-20320 is an unauthenticated XML External Entity (XXE) flaw in the Open Client Interface XML parser of Cisco BroadWorks. Cisco says a crafted XML message to the OCI-Provisioning service lets a remote attacker read files with the privileges of the BroadWorks user. NVD rates it 7.5 (CVSS), and the first fixed release is RI.2026.07.

Is CVE-2026-20320 actively exploited?

CISA added CVE-2026-20320 to its Known Exploited Vulnerabilities catalog on November 10, 2026. Malbear Labs reported that an AI agent run by a Chinese-speaking operator used it against a European telecom provider and took the XSP's system identity private key, three SSL credential sets and three more private keys. Cisco's advisory, at revision 1.2 on November 13, 2026, says Cisco has no direct evidence of active exploitation.

How do I check if my clients were hit by CVE-2026-20320?

First, preserve and hash each host's logs (check 1 in the body), then search the copies with sudo grep -rIn -E "<!DOCTYPE|<!ENTITY" /root/ir-cve-*/. Next, confirm each host's release with bwshowver and treat anything earlier than RI.2026.07 as vulnerable. Finally, run nmap -Pn -p 2208,2220 --open -iL client-public-ranges.txt from outside the client's network and close any open port it lists.

Can an autonomous pentest find exposed BroadWorks servers?

Yes. An autonomous external network pentest such as KinoSec's maps a client's internet-facing services, flags exposed ports like OCI-P port 2208 and tests them for known exploitable vulnerabilities. It then reports each finding with remediation, so you can show the client the fix worked.