N033: An Optimizely flaw let attackers plant card skimmers on 178 brands' sites
An Optimizely flaw let an attacker serve card skimmers and iOS exploit kits to visitors of 178 brands' sites, while poisoned Microsoft agent-framework releases stole cloud and CI tokens.
N033 · Wednesday 18 November 2026 · 5-min read · Coverage: 17-18 Nov
Jump to: Agent threats & vulnerabilities · Policy, protocols & compliance · Models & infrastructure · Alignment & safety · Other big security news · Pro deep dives
Today in 30 seconds
- An attacker used a flaw in Optimizely's experimentation service to write JavaScript into customer projects, and Optimizely's own CDN served card skimmers and iOS exploit kits to visitors of 178 organizations' sites.
- Two unofficial releases of Microsoft's agent-framework Python package ran a credential stealer at every Python start, taking Azure, Azure DevOps and GitHub Actions tokens before PyPI pulled them.
- Google's full report confirms attacker agents now read new patches and launch exploits with no human, so a flaw in shared vendor code gets found and used faster than teams patch.
Top story
An Optimizely flaw let an attacker plant card skimmers on 178 companies' websites
16 Nov 2026 · Report URI and Optimizely status notice · Incident · no CVE · relevance 10
A website-testing service that big brands load on every page sent their visitors malware that stole card numbers and crypto. Optimizely says an attacker exploited an unnamed flaw in Experimentation to write code into customers' Project JavaScript; cdn.optimizely.com then served skimmers, wallet swappers and iOS exploit kits for about 17 hours on sites including Zoom, BBC and HSBC. No AI role is reported. It matters now because attacker agents already hunt fresh flaws in shared code.
For leaders: Every vendor script on your pages can change your site without a release, so ask each vendor who can write to it and how fast you would know.
For builders: Search Optimizely Project JavaScript for 1688360[.]cc and atob(, rotate Optimizely credentials and tokens, and enforce a restrictive script-src CSP with violation reporting.
Agent threats & vulnerabilities
Poisoned Microsoft agent-framework releases stole cloud and CI tokens on every Python start
17 Nov 2026 · Wiz · Supply chain · agent-framework 1.26.1, 1.26.2 · no CVE · relevance 9
Two fake updates of Microsoft's Python toolkit for AI agents stole cloud and build keys from machines that installed them. Versions 1.26.1 and 1.26.2, uploaded outside Microsoft's release process, dropped a .pth file that runs at every Python start and posted environment secrets, Azure managed-identity tokens and, in 1.26.2, GitHub OIDC tokens to a webhook.
Do this: Find 1.26.1 or 1.26.2 in lockfiles, CI logs and images built since 20:35 UTC 17 Nov, pin 1.26.0 with hashes, and rotate Azure and GitHub credentials on affected hosts.
Update: Google's full report confirms attacker agents now exploit fresh patches unattended
16 Nov 2026 · Google Threat Intelligence Group · Threat report · no CVE · relevance 8
Google's own report, which would not load for N032, confirms criminals ran software that reads security fixes and attacks unpatched sites. New: GTIG calls it the first time LLM frameworks fully automated discovery and exploitation with no human. Google also sells AI defenses.
Do this: Cut patch time for internet-facing web apps to days.
Policy, protocols & compliance
New York's attorney general creates a unit to enforce the RAISE Act
17 Nov 2026 · NY Attorney General · Final · US (New York) · RAISE Act effective 1 Jan 2027 · relevance 7
New York's top law officer set up a team to check AI companies' safety promises and incident reports. The ASSET Unit will enforce the RAISE Act, which makes large frontier developers publish safety measures and promptly report security and safety incidents.
Do this: Frontier developers: name who files RAISE incident reports before 1 January.
Chart of the day
Zero-day exploits hit 22 in August, nearly three times 2025's monthly average
2025, average per month
2026, average per month
August 2026
Source: Google Threat Intelligence Group, AI Threat Tracker, published 16 Nov 2026 (primary). 2026 average as reported by GTIG.
Models & infrastructure
OpenAI's GPT-6.2 Astra beat twice as many sandbox tests as the model before it
17 Nov 2026 · OpenAI system card · Model release · closed, API and ChatGPT · Capability shift · relevance 8
OpenAI's newest work model got better at breaking out of the locked test environments meant to contain it. Its system card reports GPT-6.2 Astra solved 36.4% of Sandbox Bench sandbox-exploitation targets, up from 18.2% for GPT-6 Astra (self-reported).
New risk for leaders: The model inside your sandbox is a stronger attacker of it; isolation still works only with a second control behind it, such as egress limits and no host credentials.
Do this: Re-run sandbox escape tests with Astra before switching.
Alignment & safety
Also:
- OpenAI embeds outside testers: Apollo, METR, Transluce and Goodfire get on-site access to test for covert misbehavior (13 Nov); OpenAI can redact what they publish.
Other big security news
- Puppet Enterprise on CISA KEV: CVE-2026-115731 (CVSS 9.4 deserialization) is exploited; upgrade to 2026.0.1, 2025.12.1 or 2023.8.13 by 1 Dec (CISA) (actively exploited).
- Rakuten Securities breach: An intruder in a test system may have exposed personal and bank data for up to 4.47 million accounts (Rakuten Securities).
One concept
Content Security Policy (CSP)
A CSP is a rule your website sends browsers listing where scripts may load from and where page data may go. It still works when a trusted vendor's server delivers bad code: on 24 affected sites, a strict script-src blocked the Optimizely payload.
Basic mitigation: Name allowed hosts in script-src, drop 'unsafe-eval', limit connect-src and turn on reporting.
AgentSec Brief is published by AM Technology Solutions. Researched and drafted with AI agents, verified against primary sources, and reviewed before publication. Informational only; not legal, compliance or security advice. Verify before acting.
Pro members get the full deep dives with attack diagrams, detection queries and the fix in order, plus the vulnerability tracker, compliance calendar, open-source picks and watch list. Upgrade at https://www.agentsecbrief.com/#/portal/signup