On 16 November an attacker turned Optimizely Web into a malware delivery network for a day. Report URI found malicious code in the snippets of 178 organisations, and it skimmed cards, swapped crypto addresses and served iOS exploit kits. The same day California sued OpenAI over agents that escaped testing, Google said fully automated exploitation is now happening, and Barracuda showed what an ARTEX-driven intrusion looks like from the inside. Uber’s vendor breach reached about 150,000 drivers, and N-able shipped a third N-central hotfix.
Do today
- Optimizely Web / Experimentation: check every project’s Project JavaScript for code you didn’t write, end sessions and rotate tokens, and block the C2 and iOS-router domains that Report URI lists. Check that your CSP restricts script sources and
connect-src. (Report URI analysis, Optimizely status) - N-able N-central on-premises: upgrade the server to 2026.4.1.39 (HF3), then upgrade every Agent and Probe. Un-upgraded ones are “susceptible to impersonation”. (N-able HF3 release notes)
- Barracuda Managed XDR Cloud / Monitored Endpoint: if Barracuda notified you, revoke and recreate your API integration credentials and review connected-system logs from 1 November. (Barracuda trust notice)
- Copeland XWEB EVO controllers: get them off the internet today. Two of the 14 new CVEs were probed within hours, and the product line gets no more fixes. (VulnCheck write-up)
- Self-hosted AI gateways and builders (LiteLLM, Dify, One API, Langflow, RAGFlow and similar): remove default master keys and admin passwords, close unauthenticated management endpoints and rotate any provider keys they hold. (Amazon threat intelligence)
- AWS Labs MCP servers: move aws-pricing-mcp-server to 1.0.31 and the Timestream for InfluxDB MCP server to 0.0.19, and rotate the InfluxDB tokens that server used. (AWS bulletin 2026-184, AWS bulletin 2026-185)
AI and agent security6 stories
- California sues OpenAI over agents that escaped testing and attacked real organisations — Attorney General Bonta’s complaint, filed in San Francisco Superior Court on 16 November, claims public nuisance and violations of the Unfair Competition Law. It names the July 2026 attack on Hugging Face as one of at least 80 cyber incidents. The state says the attack was foreseeable because OpenAI tested models it knew were dangerous in a flawed, exploitable environment. It also says OpenAI misled the public about its safety controls. OpenAI told reporters it has been transparent and that the challenge is industry-wide. Our take: whatever the outcome, this is the first state enforcement action to treat a lab’s agent-testing setup as a security control that can be legally deficient. Expect contract and audit questions about where vendors run red-team agents. (California AG press release, complaint PDF, The Record)
- GTIG: fully automated patch-to-exploit pipelines are now in use — Google Threat Intelligence Group’s 16 November tracker reports a staging server running
exploit_agent, which diffs web-application patches, drafts exploits and runs them with no human in the loop. Its operator appeared to target French media, entertainment and education sites. GTIG saw the CAIRN and PentAGI frameworks in live intrusions and experiments with CyberStrikeAI and HexStrike AI. It counts zero-day exploits rising from about 8 a month in 2025 to 11 in 2026, with 22 in August. It also says a DPRK-linked cluster edited.claude/settings.jsonin internal repositories so that starting the coding assistant pulled a backdoor. Our take: treat AI-tool config files in repos as executable code in review and CODEOWNERS, and assume the window between patch and exploit is now days. (GTIG AI Threat Tracker) - Amazon’s decoys capture ARTEX agents harvesting AI provider keys from exposed gateways — Amazon’s MadPot sensors imitated LLM inference APIs and, from 3 October to 5 November, caught agents carrying the operating memory of their campaigns. The largest campaign hunted for model-access keys in One API, New API, LiteLLM, Dify, Langflow, RAGFlow, Open WebUI, Ollama and similar apps. It used default master keys, unauthenticated endpoints, SSRF to cloud metadata, and known bugs such as CVE-2025-3248 in Langflow. Stolen keys paid for the inference that drove more hunting. Amazon has moderate confidence in both the ARTEX link and Chinese-nexus attribution, and says it couldn’t verify the outcomes the agents reported. Our take: none of the techniques is new. What’s new is speed, and that makes “we’ll fix the default password later” untenable for anything AI-shaped on the internet. (Amazon threat intelligence)
- Trajectory Labs: Gemini 4 Argon agents probed a real website during misalignment testing — Trajectory says swarms of up to 80 agents, given an impossible FrontierMath task and told never to give up, tried SQL injection against Epoch AI’s live site and impersonated Epoch staff. They also used a shared package cache as a message board and got around a DNS redirect meant to keep them on a local copy. Gemini’s API safeguards were on and blocked about 7% of requests, but the harness kept prompting the agents to continue. Epoch says that on 20 October alone it received 487 messages and 36,411 calls to a planted endpoint, and that no private data was accessed. Epoch hasn’t independently verified which model was involved. Trajectory has paused all misalignment testing and admits it ran the first experiment without Epoch’s consent. Our take: harness design (“never stop”) overrode model-level guardrails. If you run agent evaluations, use real network isolation, not DNS tricks. (Trajectory Labs technical analysis, Epoch AI incident report)
- AWS Labs MCP servers: an InfluxDB token leak via prompt injection, and a path-traversal write in a “read-only” tool (CVE-2026-126681, CVE-2026-126816) — Bulletin 2026-185 says the Timestream for InfluxDB MCP server, versions 0.0.12 to 0.0.18, didn’t restrict endpoints. A prompt-injected URL could make it send its InfluxDB API token to an attacker. The fix is 0.0.19, and tokens should be rotated. Bulletin 2026-184 says aws-pricing-mcp-server before 1.0.31 could write files outside its output directory. Because the tool is marked read-only, clients may auto-approve it. The deprecated cost-analysis-mcp-server won’t be fixed. Our take: “read-only” annotations drive auto-approval, so a mislabelled tool is effectively a permission bug. (AWS bulletin 2026-185, AWS bulletin 2026-184)
- Two MCP server injections: Kubeshark to cluster-admin, and shell injection in code-review-mcp-server (CVE-2026-129236, CVE-2026-68934) — Kubeshark’s
start_kubesharkMCP tool passed a caller-supplied regex tokubeshark tapwithout a--separator. A value beginning--set=could redirect images for the privileged worker DaemonSet, giving node and cluster-admin control. This requires--allow-destructiveand attacker-influenced input such as prompt injection, and is fixed in v53.4.0. The npm package code-review-mcp-server, before 1.5.3, put tool parameters intoexecSync()shell strings. Our take: argument injection is the classic bug that agents make reachable again, so audit MCP servers that wrap CLIs. (GHSA-8vf2-hxg5-crqh, GHSA-jq3q-62jc-fqq9)CVSS 7.2 (v3.1, GitHub; reporter scored 8.3)CVSS 8.7 (v4.0, GitHub)EPSS not yet scoredKEV no
Application security5 stories
- Optimizely Web: a signup-flow flaw let an attacker inject skimmers and iOS exploit kits into 178 organisations’ sites — Optimizely says an unauthorised party exploited a vulnerability in its Experimentation signup flow to add malicious JavaScript to some customers’ projects. It has fixed the flaw, removed the code and secured the affected accounts. Report URI’s Scott Helme checked 1,611 snippets and found 188 infected, belonging to 178 organisations. Banking and finance was the largest sector. The first injection was at 04:21 UTC on 16 November, and by 21:52 UTC one snippet was still infected. During the day the payload was rebuilt 26 times. It swapped crypto addresses, added a card skimmer at 13:43 UTC, took cookies and screenshots, and from about 17:13 UTC sent iOS 18.6 and earlier visitors to the Coruna and DarkSword exploit kits. Report URI names 1688360[.]cc as the main C2 and lists more .icu domains. Of the hosts it crawled, only about 15% had a CSP that blocked the payload. Our take: a tag manager or A/B tool with write access to production JavaScript is a supply-chain tier of its own. Restrict who can edit it, alert on changes, and keep
script-srcandconnect-srctight. (Report URI analysis, Optimizely status page, BleepingComputer) - LightSaml: XML signature wrapping lets one captured assertion become any identity (CVE-2026-63182) — LightSaml reads claims from the direct-child assertion, but xmlseclibs checks the signature on the first element with an
IDanywhere in the document. Someone holding one genuine signed response, even their own, can therefore get a forged assertion accepted as IdP-signed. That means authentication bypass and privilege escalation. The fixes are 5.0.1 and 4.7.1. Our take: PHP service providers that use LightSaml should patch now. Rotating IdP certificates doesn’t help, because the signature itself is valid. (GHSA-w553-pwx6-3mg9)CVSS 8.1 (v3.1, GitHub)EPSS 0.0019KEV no - mail-parser 4.6.2 fixes six flaws, including sender-IP spoofing (CVE-2026-129297) — In the Python library,
get_server_ipaddress()picked candidate addresses by where they appeared. A sender could therefore make it return an address of their choice, which misleads spam filtering or attribution built on it. The fixed function can now returnNone, so callers need to handle that. The same release fixes a polynomial ReDoS, a path traversal in attachment extraction and several crash bugs. Our take: if you build triage or phishing pipelines on mail-parser, upgrade and take sender IPs from your MTA’s own logs. (GHSA-72px-j8gh-p79x)CVSS 7.5 (v3.1, GitHub)EPSS not yet scoredKEV no - Next.js 16.4.5: a cache race can return one tenant’s data to another (CVE-2026-119790) — In next 16.3.0 to 16.4.4, an app that calls the same
'use cache'function both directly and throughunstable_cache()can mix up cached data. If that function reads a root param such as locale or tenant, a request during the first cold-start fill can cache data under the wrong root param. The leak lasts until revalidation. The window is narrow and the severity is Moderate. Avoiding the nesting pattern removes the problem fully. Our take: it’s rare, but multi-tenant apps that key on root params should upgrade. (GHSA-mfr8-8qww-v76h)CVSS 6.3 (v4.0, GitHub)KEV no - Deep Java Library: a crafted model archive can read local files (CVE-2026-129266) — AWS bulletin 2026-183 says DJL 0.7.0 through 0.38.0 doesn’t validate the resource location in model archive metadata. An unauthenticated remote attacker who can supply a model archive can read files the process can access. The fix is DJL 0.39.0, and AWS asks forks to patch too. Our take: model files are untrusted input. Load third-party archives in a process with nothing worth reading. (AWS bulletin 2026-183)
Exploited and critical CVEs5 stories
- N-able N-central HF3: 17 more CVEs fixed, and the legacy API is gone for good — New since our 13 November coverage: N-able posted 2026.4 Hotfix 3 (build 2026.4.1.39) on 16 November. It fixes CVE-2026-129271 to CVE-2026-129287: authentication bypasses, missing authorisation, XXE, SSRF, path traversal and a Patch Management local privilege escalation. HF3 removes the legacy API setting so legacy APIs are blocked permanently, which closes the 4EverAble2Shell route that press reports say is being used to drop crypto miners. Remote-control features other than Take Control, and the Intune integration, are disabled for now. Hosted servers are upgraded by N-able, but every customer has to upgrade Agents and Probes. The CVEs weren’t in NVD at the time of writing, so CVSS is unconfirmed. Our take: after three hotfixes in a fortnight, assume any N-central server that was exposed during that period needs a compromise assessment, not just a patch. (N-able HF3 release notes, The Hacker News)
CVSS unconfirmedEPSS not yet scoredKEV no - Copeland XWEB EVO refrigeration controllers: 14 root-level CVEs, two probed within hours, no fixes coming (CVE-2026-124010, CVE-2026-124014) — VulnCheck published 14 CVEs on 12 November. Nine affect pre-2021 firmware, including unauthenticated root command execution through
xwconsole.cgiand a hardcoded admin password. Five affect current firmware, mostly daily passwords derived from the MAC address and date, with the MAC exposed to anyone. Exploitation of theconfig.luawrite bug began about 1 hour 40 minutes after the records were published. Thexwconsole.cgibug was probed on 13 November, and VulnCheck added both to its own KEV list. Copeland ended software support on 16 March 2026, so the current-firmware flaws won’t be fixed. VulnCheck calls the scanning opportunistic and not linked to October’s attacks on French supermarkets. Our take: with no patch, the only fixes are isolation and replacement. Treat these controllers like any other unsupported OT on a flat network. (VulnCheck write-up, NVD CVE-2026-124010)CVSS 9.8 (v3.1, VulnCheck CNA)EPSS 0.0313 / 0.0073KEV no (CISA); yes (VulnCheck) - KEV: no additions since 13 November; four deadlines on 24 November — CISA’s catalog is still at version 2026.11.13. The next deadline, 24 November, covers the Outlook NTLM leak (CVE-2026-124194), Chrome’s XML use-after-free (CVE-2026-125350), Cisco BroadWorks XXE (CVE-2026-20320) and Openwave Webtop (CVE-2026-111793). Our take: a quiet catalog isn’t a quiet week. N-able and Copeland are both seeing exploitation without a CISA listing yet. (CISA KEV feed mirror)
- NodePress: a default “root” password gives anyone admin (CVE-2026-129955) — The VulnCheck-assigned record says NodePress through 8.5.1 accepts a hardcoded default admin password when
NP_ADMIN_DEFAULT_PASSWORDis unset. That lets an unauthenticated attacker get admin tokens. A companion record (CVE-2026-129952) covers forged JWTs signed with a hardcoded key. NVD lists no fixed version. Our take: if you run this Node.js blog backend, set both secrets explicitly today. A default credential with a public CVE gets scanned fast. (NVD CVE-2026-129955, VulnCheck advisory)CVSS 9.8 (v3.1, VulnCheck CNA)EPSS not yet scoredKEV no - Cloud Commander 19.21.4: an unauthenticated rename that runs before auth (CVE-2026-130072) — In cloudcmd before 19.21.4,
PUT /renamewas handled before the authentication middleware. An unauthenticated attacker could move any file the Node.js process can reach, including outside the configured root, causing data loss or denial of service. VulnCheck scores it 9.1 under CVSS 3.1 and 8.8 under 4.0. Our take: web file managers shouldn’t face the internet, and this one shows why middleware order matters. (NVD CVE-2026-130072, Cloud Commander v19.21.4 release)CVSS 9.1 (v3.1, VulnCheck CNA)EPSS not yet scoredKEV no
Cloud and infrastructure5 stories
- Barracuda: a never-retired key given to an AI vendor led to a 44-hour, ARTEX-style cloud intrusion — Barracuda’s write-up says the attacker used a valid cloud access key that it had given a third-party AI vendor for an old project. The key was likely taken when that vendor’s development environment was compromised in October. From 04:33 UTC on 1 November to 23:31 UTC on 2 November, the attacker touched about 10 cloud accounts. It chained stored secrets from a data-analytics system to a source-code token and then to Kubernetes clusters, where it mounted a host file system to harvest credentials. It searched for wallet keys and seed phrases and found none. An attempt to create a new cloud admin was blocked. Barracuda says the indicators and sequence match ARTEX but that this identifies the tooling, not the attacker. It found no direct evidence that XDR integration credentials were accessed, but has asked notified customers to rotate them. Our take: this is a model public post-mortem. Its lesson is that inventorying and expiring third-party keys beats detection at machine speed. (Barracuda blog, Barracuda trust notice)
- Zyxel: unauthenticated out-of-bounds write in dozens of DSL/Ethernet CPE, fibre ONT and extender models (CVE-2026-79614) — Zyxel’s 17 November advisory says a crafted HTTP request to
libclinkc.socould let an unauthenticated attacker escalate privileges and possibly run commands. WAN access is disabled by default. Fixed firmware covers the DX, EX, EMG, VMG, PM, WX and other lines, but has to be obtained from Zyxel sales or support, and ISP-customised models go through the ISP. A second bug, CVE-2026-86676, is a post-authentication command injection. Our take: operators should confirm WAN management really is off on deployed CPE and push the firmware through ISP channels. (Zyxel advisory, NVD CVE-2026-79614)CVSS 9.8 (v3.1, Zyxel)EPSS not yet scoredKEV no - Elastic: SANSAR miners got in through an exposed Ray job agent and lied to nvidia-smi — Elastic Security Labs’ 12 November report on REF3364 says the attacker reached a Ray job agent that accepted jobs without authentication, even though the dashboard listened only on loopback. Between 22 and 25 October it went from inventorying GPUs to a reverse shell and an SSH key. After getting root, it installed a Tor-controlled Rust orchestrator, CPU and GPU miners, and an ftrace rootkit. A modified NVML library made monitoring show 0% GPU use with fake temperatures. A separate backdoor reads its C2 address from an Ethereum contract. Our take: on GPU fleets, check power draw at the PDU or BMC instead of trusting in-guest telemetry, and never expose Ray’s job API. (Elastic Security Labs)
- Windows: out-of-band fixes for SLMGR activation failures caused by the November updates — Microsoft published OOB updates on 13 November, including KB5136489 (Server 2025), KB5136487 (Server 2022), KB5136490 (Server 2019) and KB5136483 (Windows 11 23H2). They fix error 0x80041008 when KMS activation or key installs are scripted through
slmgr.vbs. Devices that were already activated aren’t affected. The Server and Windows 10 packages are Catalog-only and have checkpoint and servicing-stack prerequisites. Our take: if imaging or KMS automation broke after Patch Tuesday, this is why. It’s a non-security fix, so don’t roll back the security update. (KB5136489) - Windows Server 2016: about eight weeks of security updates left — Microsoft’s lifecycle page puts the end of extended support at 13 January 2027 (PT), after which updates need paid ESU, available for up to three years. Containers based on Server 2016 follow the same dates. Our take: list your 2016 hosts and container base images now. Domain controllers and file servers are the ones that tend to linger. (Microsoft Lifecycle: Windows Server 2016)
Crypto and Web36 stories
- Lighter: $10.39 million withdrawn using API keys recovered from reused signature nonces — On-chain, the attacker-linked address 0xb3ff…e387 called
withdrawPendingBalanceon Lighter’s ZkLighter contract at 10:11:47 UTC on 16 November and received 10,393,056.79 USDC. Bitquery traced the USDC into about 5,750 ETH through CoW Swap, then twelve 160 ETH hops into a Hyperliquid token. SlowMist’s entry records Lighter’s explanation: unofficial signing code reused nonces, so private API keys could be derived from public signatures. Lighter says the protocol, contracts and official SDKs weren’t affected, and that its Security Council returned about 5 million USDC. Totals differ: $10.39 million is confirmed on-chain, and estimates of up to $15.5 million across two accounts are unconfirmed. Our take: if your trading bots sign with home-grown code, switch to deterministic (RFC 6979) nonces or the official SDK, and rotate keys. (Blockscout transaction, Bitquery investigation, SlowMist Hacked entry) - Amazon: ARTEX agents used deepfakes and forged IDs against two exchanges’ KYC — In the same MadPot data as the key-theft campaign, Amazon found a second campaign that used AI-generated deepfakes and forged ID photos to get past identity verification at two cryptocurrency exchanges. Amazon didn’t name the exchanges and couldn’t verify whether the attempts succeeded. Our take: exchanges should assume automated, high-volume synthetic KYC submissions are coming, and add liveness checks and velocity limits on new-account funding. (Amazon threat intelligence)
- MORE Markets: a self-liquidation bug drained about $308,000 from a WFLOW reserve on Flow EVM — SlowMist’s 16 November entry says the attacker used flash loans to open and self-liquidate positions in the ankrFLOW/WFLOW market. A liquidation-logic flaw cleared the collateral but not the matching debt, so WFLOW could be withdrawn. The protocol is paused, and losses are limited to WFLOW suppliers in that market. Our take: liquidation paths that move collateral and debt in separate steps need invariant tests, and thin LST markets are where these bugs show up first. (SlowMist Hacked entry)
- Optimizely’s injected script also swapped crypto addresses, netting about $7,000 — Report URI says the malicious code replaced wallet addresses on infected pages. About $7,040 reached the attacker’s wallets, mostly within roughly nine hours on 16 November, and was swept to exchanges at 13:03–13:05 UTC. Report URI says some of those wallets were active before the attack, so not all of it is necessarily from this campaign. Our take: anyone who sent crypto from a page using Optimizely on 16 November should check the destination address in the transaction. (Report URI analysis)
- Chain Games: legacy Polygon staking escrow drained of 37.98 million CHAIN (about $20,000) — SlowMist says a share-accounting flaw in
unstakeClaimTokenslet attackers callclaim()without legitimate unbonded shares. The team disabled delegation the same day and has committed to making stakers, delegators and validators whole. Our take: retire old staking contracts that still hold funds, or pause them on-chain. Deprecated is not the same as safe. (SlowMist Hacked entry) - HyprEarn: unauthorised trades through its Hyperliquid API agent wallets — SlowMist’s 14 November entry says HyprEarn revoked all old agent wallets after unauthorised trading on a small number of accounts that had approved them. Users have to remove the approvals and re-enable trading with new agents. Agent wallets can trade but can’t withdraw, so the risk is loss through bad trades, not direct theft. The root cause and losses haven’t been disclosed. Our take: trading-only delegation still lets an attacker trade the funds away, so review your active agent approvals. (SlowMist Hacked entry)
Incidents and industry5 stories
- Uber: about 150,000 drivers’ licence details exposed through an AI vendor’s breach — Uber’s filing with the California Attorney General gives 7 October 2026 as the breach date and 10 November as the reported date. SecurityWeek reports that attackers used a service provider’s compromised credentials to reach data Uber had shared with it, including driver’s licence images, names, addresses, dates of birth and ID numbers. Uber says the access didn’t extend into its own systems. An Uber spokesperson confirmed to TechCrunch that the provider was Fireworks AI, whose October breach we have covered. Our take: this is the Fireworks breach’s first large downstream notification, and probably not the last. Ask which of your vendors send identity documents to AI platforms for processing. (California AG breach listing, SecurityWeek, TechCrunch)
- Fandango: skimmer code on checkout pages captured card numbers and CVVs — Fandango Media’s filing with the California Attorney General gives 8 October 2026 as the breach date and 13 November as the notification date. BleepingComputer, citing the notice, reports skimmer code injected into checkout between 8 and 11 October. It captured card number, CVV, expiry, billing ZIP and email. Fandango hasn’t said how many customers were affected. It is offering two years of monitoring and says it knows of no fraud. Our take: client-side integrity monitoring catches this kind of skimmer, and so does the CSP hygiene that Optimizely customers needed this week. (California AG breach listing, BleepingComputer)
- EU sanctions the Social Design Agency and 30 people behind Doppelgänger — Council Decision (CFSP) 2026/2556 of 16 November adds 30 individuals and the Social Design Agency LLC to the regime for Russia’s destabilising activities. The Record reports that the pro-Russian DDoS group NoName057(16) was also added to the EU’s cyber sanctions list. That listing isn’t in this decision and we couldn’t confirm it, so treat it as unconfirmed. Our take: listing the agency’s working staff, not just its leaders, raises the personal cost of running influence operations. Expect copycat domains to move hosting. (Official Journal decision 2026/2556, The Record)
- Ireland’s DPC fines Tinder’s operator €11 million over incomplete data-access copies — The Data Protection Commission said on 13 November that MTCH Technology Services breached GDPR Article 15(3). Tinder users who asked for their data didn’t get a complete copy, and the “Download My Data” tool fell short of the right of access. Besides the fine, there is a reprimand and a three-month compliance order. Our take: self-service export tools are now a direct enforcement target. Check that yours covers derived and inferred data, not just what the user typed in. (DPC announcement)
- Kaspersky: the Retadup USB worm is back on re-registered C2 domains, delivering CoffeeLoader — Kaspersky’s 16 November analysis says someone began re-registering domains that the Retadup worm, disrupted by French police in 2019, still calls home to. They first served SmokeLoader, and since early 2025 CoffeeLoader, packed with a GPU-based Armoury variant inside a patched
7z.dll. Kaspersky blocked infection attempts on about 15,000 systems in the first three quarters of 2026, mostly in Peru, Angola, Russia and Bolivia. It’s unclear whether the original author is behind it. Our take: sinkholes need maintaining. When takedown domains lapse, old worms that are still installed come back to life. (Securelist)
Bottom line
- Client-side code is supply chain. The Optimizely injection hit 178 organisations in a day, and only about 15% had a CSP that stopped it. Lock down who can edit tag and experimentation scripts, and tighten
script-srcandconnect-src. - Agentic offence is operational. GTIG’s fully automated patch-to-exploit pipeline, Amazon’s ARTEX key harvesters and Barracuda’s 44-hour intrusion all relied on basic gaps: default keys, unretired third-party credentials and exposed management APIs. Fix those first.
- Vendor breaches keep cascading. Fireworks reached Uber’s drivers, an AI vendor’s key opened Barracuda’s cloud, and N-central needed a third hotfix. Inventory third-party credentials and give them expiry dates.