CISO Tribune

Town Square · Tuesday, November 17, 2026

Optimizely code injection hit major brand sites and 4 more

By Hadass Liza Bitton · November 17, 2026 · 5 stories

Optimizely confirms code injection that put malicious scripts on major brand sites, and the claim holds: the company's own notice confirms a signup-flow flaw let an attacker add code to customer projects reaching visitors of Zoom, GitLab and the BBC, which Report URI and Hudson Rock documented independently. Two other threads ran ahead of the evidence, since the Chase software bug that slowed online banking at several US banks was a bug, not the nation-state attack Reddit threads alleged, and Google's report that artificial intelligence (AI) attackers now automate exploitation stands on its own without the unverified Reddit claim that a July Hugging Face incident proves it. In our view, the California attorney general's suit against OpenAI is real but its allegations still await a court, and VulnCheck's count of 37,966 Glasswing findings against 217 fixes is documented, while its claim that maintainers broadly are overwhelmed reaches past the curl evidence.

Today's board

  1. 1Optimizely confirms code injection that put malicious scripts on major brand sitesConfirmed
  2. 2Chase software bug slowed online banking at several US banks; bank says no cyberattackConfirmed
  3. 3California attorney general sues OpenAI over safety claims and Hugging Face attackConfirmed
  4. 4Google says AI attackers now automate exploitation and projects full autonomy within monthsConfirmed
  5. 5VulnCheck says Anthropic's Glasswing has 37,966 findings but only 217 fixesPartly confirmed

Story 1

Optimizely confirms code injection that put malicious scripts on major brand sites

A flaw in Optimizely's signup flow let an attacker add code to customer projects, reaching visitors of Zoom, GitLab, the BBC and others.

Our verdictConfirmed

Optimizely's own incident notice, as reported by InfoStealers, confirms the vulnerability and the project-code injection, and Report URI and Hudson Rock documented the malicious code independently.

What people are saying

Report URI said that on November 16, 2026, malicious JavaScript was served to visitors of 178 organizations through the Optimizely Web experimentation platform. Its list includes Zoom, the BBC, CNN, Salesforce, GitLab and HSBC. InfoStealers reported that the confirmed compromise is malicious code served from third-party providers' own script and widget endpoints.

What holds up

Optimizely's official incident notice, posted at 01:02 UTC on November 17, confirms an Experimentation vulnerability and malicious project-code injection, InfoStealers reported. According to a follow-up from InfoStealers, Optimizely told BleepingComputer that an unauthorized party exploited a vulnerability in its Experimentation account signup flow and added malicious JavaScript to customer projects. Optimizely says it patched the vulnerable path, removed injected code, ended sessions, revoked tokens and reset access for affected accounts.

Report URI found the code in Project JavaScript, a field of custom customer code that runs on every page. It said the kit swapped cryptocurrency wallet addresses, skimmed payment cards, stole cookies and screenshots, and delivered the Coruna and DarkSword iOS exploit kits to iPhones and iPads.

Hudson Rock independently matched the archived hashes and directly recovered nine poisoned vendor files. The same work found poisoned responses from Kissmetrics, StorifyMe and minute-ly, and says each provider's initial access mechanism remains unknown. The evidence reviewed does not establish content delivery network (CDN) cache poisoning or breaches of every affected brand's own servers. Report URI said that late on November 16 it found the code in only 1 of 188 infected snippets.

We rate the core claim confirmed: the compromise rests on the vendor's own notice and two independent technical analyses, while attribution and full scale remain open.

Inside the threads

In a Hacker News thread, one commenter faulted tag tools that let marketing teams add code to a site without involving engineering, and another called client-side A/B testing services a security nightmare. One commenter urged Subresource Integrity (SRI), and a reply suggested it goes unused because vendors continuously update their scripts.

Our take

We think the event is real and the vendor fix is the easy part. In our reading, the harder issue is that a field marketers control can push code onto every page of a corporate site with no security review. Our take is that crypto losses are the visible damage, and stolen session cookies are the quieter one.

What to do

  • Ask web and marketing teams which pages load Optimizely Web or other third-party tags, and who can publish code through those tools.
  • Report URI says a CSP in enforcement mode could have blocked key parts of the attack; we would move payment and login pages to enforcement first.
What the verdict rests on (3 sources)

Story 2

Chase software bug slowed online banking at several US banks; bank says no cyberattack

JPMorgan Chase blamed an on-premise hardware software bug. Reddit threads pushed a nation-state theory that no source supports.

Our verdictConfirmed

AP, the BBC and Al Jazeera each reported the disruption, and a JPMorgan Chase spokesman confirmed it to AP as a software bug that is not cyber-related.

What people are saying

What holds up

The Associated Press (AP), in a report carried by PBS, said thousands of customers had trouble reaching online accounts at several major banks and at Zelle after a technical issue at JPMorgan Chase caused ripple effects throughout the financial system. Bank spokesman Joe Evangelisti said Chase had "a temporary issue related to a software bug in an on-premise hardware system" and that "it is not cyber-related."

He also said "This is not a systemic issue," adding that other banks may have been affected by slower transactions only to the extent they were a party to a transaction with Chase. AP reported that other banks trying to transact with Chase, including Bank of America and Citigroup, were not having issues with their banking services. A spokesman for Early Warning Services, which owns Zelle, told AP that the Zelle network was fully operational.

The BBC said reports on Downdetector began around noon eastern time and later dropped considerably. Al Jazeera reported nearly 5,000 issue reports at Chase and more than 800 at Wells Fargo.

The disruption is confirmed. The nation-state attack theory has no support in any source, and the bank at the center of it says the cause was not cyber-related.

Inside the threads

The loudest camp tied the outage to a run of recent failures and read it as covert conflict; a skeptic replied that the war is "With tech debt and complexity, it seems."

Our take

We think the attack theory is overblown, since JPMorgan Chase said on X that a software issue caused the outage and that there is no evidence it is cyber-related, Al Jazeera reported. United Airlines, which had its own outage Monday, said there was "no indication that this was a cybersecurity-related issue," Al Jazeera reported. What is underrated is the concentration risk. In our reading, one bank's internal fault slowing counterparties and Zelle payments is a third-party dependency story, and boards should hear it that way. We also think security teams carry a communications job here: confident misattribution inside your own company wastes incident response time. Treat early cyber speculation as a hypothesis to test, not a finding.

What to do

  • Identify which payroll, treasury and customer payment flows depend on one banking partner, and confirm a manual fallback exists.
  • Brief executives that the bank attributed this to a software bug, so rumors do not trigger unneeded incident escalation.
  • Add correlated provider outages to tabletop exercises, including how you will confirm a vendor's cause before calling it an attack.
What the verdict rests on (3 sources)

Story 3

California attorney general sues OpenAI over safety claims and Hugging Face attack

The state's complaint is real and on file. Its allegations about OpenAI's testing controls are claims a court has yet to weigh.

Our verdictConfirmed

The California Attorney General's own press release confirms the lawsuit, its legal theories and the quoted statement tying OpenAI's internal culture to the Hugging Face attack.

What people are saying

A Register report says California's attorney general sued OpenAI under unfair competition and public nuisance laws for allegedly misleading and endangering the public. The Register says the attorney general cites the Hugging Face incident as an example of why the industry must be accountable.

What holds up

The lawsuit exists. The attorney general's press release says the suit alleges OpenAI misled and endangered the public by testing advanced artificial intelligence (AI) models without adequate safeguards and monitoring. The release says the case was filed in San Francisco Superior Court. It alleges both a public nuisance and a violation of California's Unfair Competition Law.

The quote at the center of the chatter is accurate. Bonta said the attack was the predictable result of an internal culture that put growth ahead of safety. He called it one of 80 known incidents. The release says the attack lasted over two days and involved coordinated actions by hundreds of autonomous AI agents.

According to NBC News, an OpenAI spokesperson said the suit appears to single out the company over a challenge every major lab has identified.

What is not established is liability. The core claim is confirmed, but the complaint's allegations remain allegations.

Inside the threads

On Mastodon, one commenter said the case looked fast-tracked for a public prosecution. The same commenter asked why the state had not used anti-hacking laws. Another commenter wondered whether computer misuse laws could reach general AI scraping. The thread then drifted into speculation about the Supreme Court that no source here supports. Another Bluesky post said staff from outside research groups investigated the incident.

Our take

In our reading, the headline risk here is not rogue AI. It is the gap between what a vendor says publicly about its controls and what it does in practice. The state's theory treats safety representations as consumer claims that can be tested under unfair competition law. We think security leaders should expect that framing to reach vendor contracts and questionnaires. The rogue AI storyline is overblown. Katharine Trendacosta of the Electronic Frontier Foundation told The Register that several recent incidents at AI labs could have been prevented or mitigated by long-standing practices such as sandboxing, monitoring and logging. Our take: any AI vendor you rely on should be able to explain the controls around its testing as clearly as the controls around its product.

What to do

  • Add questions to AI vendor due diligence about how test and research environments are isolated from the internet and from customer systems.
  • Ask vendors for their incident history involving autonomous agents, and how they notified affected parties.
  • Apply sandboxing, monitoring and logging to any agents you deploy internally, scoped by how reversible and sensitive each action is.
  • Have legal review any public or contractual statements your own organization makes about AI safety controls.
What the verdict rests on (4 sources)

Story 4

Google says AI attackers now automate exploitation and projects full autonomy within months

Google's threat team reports fully automated exploitation already in use. Reddit's claim that a July Hugging Face incident proves it is unverified.

Our verdictConfirmed

Google Threat Intelligence Group's own AI Threat Tracker report states the findings that BleepingComputer summarized.

What people are saying

BleepingComputer reported that Google warned fully automated AI attacks are coming soon, and the report was discussed on Reddit's r/cybersecurity.

What holds up

Google's threat report says attackers are deploying autonomous multi-agent offensive frameworks such as ARTEX, CAIRN, CyberStrikeAI and PentAGI. The same Google report says attackers use large language models (LLMs) to reverse engineer patches and weaponize known flaws against internet-facing servers without human intervention.

The headline undersells part of the report. According to BleepingComputer, Google said that in the past quarter it saw attackers fully automate vulnerability discovery and exploitation without a human in the loop for the first time. BleepingComputer reported that hackers used one such multi-agent system, called exploit_agent, against French media, entertainment and education websites. Google's own summary post is titled "The era of autonomous attacks has arrived."

The "coming soon" part is a projection. BleepingComputer reported that the researchers expect unsophisticated attacks relying mostly on AI agents within three months, and sophisticated, fully autonomous ones within six. BleepingComputer also cited a Federal Bureau of Investigation (FBI) advisory saying multiple threat actors have weaponized the ARTEX framework since August.

The verdict: confirmed that Google said this.

Inside the threads

Our take

We think the useful signal is the patch-diffing pipeline, not the six-month forecast. If attackers can read a vendor update and produce a working exploit without a person involved, your patch window for internet-facing systems is now set by machines. That makes exposure management a timing problem, not only a coverage problem. The Hugging Face debate is a distraction until someone with evidence describes what happened. The underrated risk is noise: AI-generated fake exploit code can burn analyst hours just when speed matters most.

What to do

  • Set and measure a short patch window for internet-facing applications, and treat n-day flaws on exposed hosts as emergencies.
  • Inventory and lock down the credentials your teams use for AI services. Google says financially motivated attackers steal LLM credentials and resell access to AI models through underground proxy services.
  • Build a triage step for public proof-of-concept code. BleepingComputer reported that Google's researchers cited PHP2Shell, where hundreds of purported proof-of-concept exploits, mostly AI-generated, appeared shortly after PHP announced two flaws.
  • Ask your detection provider how its automation and human analysts split the work, and how fast a new exploit reaches a working detection.
What the verdict rests on (3 sources)

Story 5

VulnCheck says Anthropic's Glasswing has 37,966 findings but only 217 fixes

A VulnCheck review of Anthropic's disclosure ledger argues the triage work has shifted to open source maintainers, with curl as the test case.

Our verdictPartly confirmed

VulnCheck's November update documents the ledger figures and the curl tally, but the claim that maintainers broadly are overwhelmed goes beyond that evidence.

Where it was argued

What people are saying

LWN covered a VulnCheck blog post that reviews what Anthropic's Project Glasswing has produced since launch. VulnCheck argues, among other points, that the triage bottleneck has moved from Anthropic to the maintainers receiving the findings.

What holds up

The figures come from VulnCheck's November update, which tracks Anthropic's own ledger alongside CVE records. Of the 37,966 findings Anthropic reports, 7,349 are public and 217 are marked fixed, according to VulnCheck. About 1,654 have been marked duplicates, nearly eight times the number fixed, and 419 have been withdrawn, VulnCheck says.

Since September, Anthropic added nearly 12,000 findings and recorded 15 more fixes, by VulnCheck's count.

VulnCheck reported that curl maintainer Daniel Stenberg announced the project had received a batch of 45 reports from Anthropic. Anthropic rated one as medium severity and the rest as low, the post says. So far 34 have been published, and curl marked 19 as not a vulnerability, confirmed six as low severity and left nine awaiting assessment, per VulnCheck. VulnCheck quotes Anthropic describing its OSS Scanner reports as "fully model-generated, without human review or triage."

In our reading, the ledger figures and the curl tally hold up, while the wider claim that maintainers in general are overwhelmed goes further than they do. We rate it partly confirmed.

Inside the threads

VulnCheck said the jump in duplicates is what it warned about in September: the longer findings sit, the more likely someone else finds and fixes them first. The VulnCheck post also says readers shouldn't overlook that AI is an effective tool for finding vulnerabilities. A bug on Anthropic's ledger appears to inflate the count of findings marked disclosed, according to VulnCheck, so we treat the disclosure figures as provisional.

Our take

We think the 217 fixes matter more than the 37,966 findings, the two counts VulnCheck took from Anthropic's ledger as of its November 2 update. In our view, a finding that sits unfixed and later turns into a duplicate costs a maintainer time and gives defenders nothing. The danger story looks overblown to us for now, since VulnCheck counts only two of the 442 CVEs it attributes to Anthropic as exploited in the wild. What we see as underrated is the effect on vulnerability data. In our reading, if low-severity, model-generated reports keep becoming CVEs, prioritization built on CVE counts or raw severity gets noisier. We think your real exposure sits in open source components whose maintainers are spending triage time on reports that often turn out not to be real.

What to do

  • We suggest making sure your software composition tooling drops withdrawn or rejected CVEs from remediation queues.
  • For open source projects your business depends on, we suggest checking whether maintainers have the funding and people to triage incoming reports, and supporting them where you can.
What the verdict rests on (2 sources)

How Town Square works

Every day we read what the security community is discussing on Hacker News, Reddit, Mastodon and Lobsters, and what search shows of the conversation on X. For each story we check the claim against primary sources, give a verdict, digest what the threads said, and say what we think.

What each verdict means

The CISO Tribune briefing

Every verified CISO move and one piece of analysis, weekly.

Unsubscribe any time. See the privacy policy.