The Vendor Was Trusted. The Access Was Too Broad.
This week gave us a useful correction to the way we talk about technology risk. The dangerous thing wasn’t always a bug in the product a business bought. Sometimes it was a perfectly valid credential held by a vendor, an extension that an administrator installed on purpose, or a routine network change performed by the company responsible for the network.
That sounds less dramatic than “hackers broke in.” It’s also more useful. If the trusted path is the attack path, the work isn’t finished when somebody changes a password or installs a patch. You have to know what that trust could reach, what it could copy, what it could change, and how you would notice when it started doing something it was never meant to do.
The Big Story: One Vendor Breach Became Four Different Incidents
Fireworks AI provides the infrastructure companies use to run, evaluate, and train AI models. In October, attackers obtained Fireworks credentials. Fireworks initially said the known impact involved secrets stored for evaluation jobs. Its November 9 update expanded that picture to include account and billing information and, for a limited set of customers, secrets, datasets, and incomplete model weights.
Then the customer disclosures started arriving.
Notion found AI chat transcripts from its Insights feature in the exposed data. Uber said a technology provider’s stolen credentials were used to reach driver information, including images of licenses and the names, addresses, birth dates, and license numbers extracted from them. Cursor said an access key it had issued to Fireworks let the attacker copy historical coding-agent traces from a read-only storage bucket.
Those are three different data types with three different consequences. A Notion transcript might contain a contract detail, customer name, password, or API key somebody pasted into a conversation. A driver’s license can support identity fraud. A coding trace can contain credentials with access to the next system. “Fireworks was breached” is factually true and operationally incomplete. Each customer created its own blast radius through the data and authority it gave the provider.
Read-only access can still open the next door
Cursor’s incident is the clearest demonstration. The Fireworks key only provided read access to a storage bucket. That sounds reassuring because the attacker couldn’t alter the bucket. But some of the stored coding traces contained credentials that secret filtering had missed. Cursor says fewer than 0.01% of customers had API credentials exposed. That is a small percentage. It is also beside the point if your credential is the one in the sample.
The attacker used credentials found in those traces to reach selected Cursor services, archived nonproduction source code, support records, usage data, conversation summaries, and other internal material. One stolen credential then allowed malicious copies of Anthropic’s Claude Code extension to be placed directly into the cache serving Cursor’s extension marketplace.
At most 5,000 users who requested versions 2.1.312 or 2.1.313 during an 11-hour window may have received the malicious build. The code tried to install an attacker-controlled SSH key, enable remote access, create a reverse tunnel, and add persistence. Some of that probably failed because of coding mistakes or missing administrator rights. “The malware may have been bad at its job” is better news than the alternative, but it isn’t a cleanup procedure. Cursor’s guidance says uninstalling the extension alone may not remove everything it dropped.
Technical discovery & auditing
The public page doesn’t tell you much about the machinery behind it. Raymond Tec audits inherited and long-running projects to uncover the plugins, integrations, data, dependencies, and old decisions that determine what the next change will really involve.
The data inside the conversation still identifies you
Notion says the transcripts for most affected workspaces were “unattributable,” meaning they weren’t labeled with a user, company, or workspace. That limits one kind of exposure. It doesn’t remove the information people typed into the conversation. Notion revoked exposed Notion API keys and reported other exposed third-party keys to their issuing vendors. An unlabeled transcript containing a customer name, internal hostname, or live credential can identify itself quite effectively.
Uber’s case has a different shape. Fireworks had legitimate access to cloud storage created for its work. The hard question wasn’t whether the vendor could log in. It was supposed to log in. The useful questions are whether it needed images of driver licenses, whether it needed the extracted text, how long that information had to remain available, and whether a valid credential retrieving an unusual amount of it would trigger an alarm.
This is where third-party access security becomes more than a vendor questionnaire. A SOC 2 report can tell you something about a provider’s controls. Fireworks lists several security certifications. None of that answers the specific question, “What can the credential we issued to this provider read today?” Certifications matter. Architecture matters more during the incident.
A worked example for a small business
Suppose a small online store connects an AI support service to its help desk. The service needs recent ticket text to draft replies. It probably does not need every ticket since 2019, raw payment exports, administrator notes, or a permanent token that can download the entire customer table. If the integration can only read new support tickets, can’t retrieve payment data, and loses access automatically when the contract ends, a vendor breach is still serious. It is not automatically a company-wide data breach.
Now change the assumptions. The same integration uses one broad administrator token, retains full conversations for model evaluation, and can reach the file-storage account where employees keep exports. The provider still performs the same visible job. The hidden authority is completely different.
I don’t think the practical response is to ban every outside AI provider. That would be easy advice to give and nearly impossible advice to operate. The response is to inventory which vendor identities exist, narrow each one to the data and actions the job actually requires, separate production from testing, set expiration dates where the platform supports them, and remove access when the relationship changes. Then test whether logs can answer who retrieved what.
Also, stop treating “don’t paste secrets into chat” as the entire control. Employees will make mistakes. Filtering should catch obvious credentials, retention should be limited, and an exposed key should have a small enough scope that rotating it is an incident task rather than an organizational archeological dig.
Turning on AI is the easy part
Deciding what an AI tool should be allowed to see, who should use it, what work it should perform, and what happens when it gets something wrong is the more interesting problem. Raymond Tec helps businesses connect and automate the tools they actually use without treating every new feature like a button that obviously needs to be switched on.
Remote Support Tools Became Remote Attack Tools
The week’s N-central and ScreenConnect incidents are related by function, not by a shared attacker. Both products are control planes. Managed service providers use them to monitor systems, run commands, install software, and provide remote support across many customer machines. That trusted reach is the product. It is also the reason an attacker values the server more than any single endpoint.
N-able took hosted N-central systems offline after a researcher published more than 30 alleged Able2Shell vulnerabilities before a fix was ready. N-able released Hotfix 2, restored hosted systems, and required a careful Agent-first, Probe-second rollout. Then new exploit code appeared. The 4EverAble2Shell chain worked against Hotfix 2 when the nondefault Legacy API was enabled and exposed to the internet. Independent researchers reported compromises and cryptocurrency miners on production servers.
That does not mean Hotfix 2 failed to do anything. It fixed the flaws it was built to fix. A patch is not a force field against the next disclosure. Operators still need HF2, should confirm the Legacy API is disabled or tightly restricted, and need to investigate systems that exposed it. Disabling the API closes the described entrance. It does not evict somebody who already obtained root.
The ScreenConnect campaign kept moving
ScreenConnect attackers abused a third-party extension called RMM+. Versions 1.0.31 and 1.0.32 could accept a blank access key when an administrator had never configured one. The attackers used legitimate Run commands to install a second, attacker-controlled ScreenConnect client and a toolkit called SCKit on managed endpoints.
Huntress initially observed at least 288 affected endpoints across 48 organizations. On November 13 it saw another wave affecting 205 hosts across 41 organizations. SCKit also tried to trap technician session cookies and spread from an infected customer machine to the technician’s workstation. Huntress and ConnectWise say they haven’t observed that spreading technique succeed. ScreenConnect technician clients version 26.6.5 and later block the client flaw SCKit tried to use, but updating the client and disabling the vulnerable extensions address separate problems. Both matter.
If your business receives remote support, don’t delete the support agent and hope for the best. Ask the provider whether it uses N-central or ScreenConnect, whether the relevant API or extensions were exposed, and whether it reviewed command history and unexpected remote-client installations. If you operate the tools, preserve the evidence and investigate through a separate security channel. Using the possibly compromised remote-support path to inspect the compromise is how one workstation problem becomes a technician problem too.
The boring machines need attention too
Browsers, workstations, remote-access tools, Wi-Fi, and ordinary office hardware rarely get much attention until one of them becomes the problem. Raymond Tec provides onsite IT and field services around Reading, Pennsylvania, along with practical help keeping the technology people use every day working and reasonably secure.
The FAA Outage Was a Change-Control Failure, Not a Cyberattack
The FAA halted departing flights nationwide for more than four hours Friday after communications and radar systems became unstable. Officials say there is no evidence of a cyberattack. A contractor was replacing an old copper telecommunications line with fiber and installing a router at a remote location. Either the router was faulty or the installation was wrong. The new device began flooding other routers with more than 1,000 connection attempts every four or five seconds.
That was enough to disrupt the systems controllers use to communicate with aircraft and receive radar information. About 3,500 aircraft already in the air were landed safely. More than 8,000 flight operations resumed after the stop lifted, although delays and cancellations continued through Saturday.
The boring explanation is the important explanation. Routine work on a legacy network caused nationwide consequences because the failure propagated beyond the change site. Businesses run smaller systems, but the principle scales down perfectly: test against realistic load, define the rollback before the change, and make sure the rollback does not depend on the system that is failing.
Microsoft 365 had its own global outage this week after an earlier capacity upgrade and a later network change combined in a way that degraded Exchange Online and dependent services. You don’t need a second complete cloud suite for every business. You do need current contact information outside the Microsoft tenant and an agreed place to post instructions when email, Teams, identity, and the status dashboard all become part of the same problem.
A campaign called wfx-loot used 4,837 compromised GitHub accounts to probe 61,764 repositories beginning November 10. Aikido Security found evidence that the malicious workflow ran in at least 1,835 repositories. The payload collected GitHub Actions secrets, cloud credentials, SSH keys, deployment tokens, payment-provider keys, AI service credentials, and other material available to the automated build environment.
The attacker then reverted the code changes and deleted the workflow run. A missing run number may be one of the few remaining clues. More importantly, rotating the obvious GitHub token may not finish the cleanup. The malware could alter AWS trust policies or add Azure federated identities, creating a new route back in that survives the first credential reset.
This was not a new GitHub Actions vulnerability. It was abuse of legitimately issued credentials that had been stolen. That distinction keeps appearing this week because identity is increasingly the control surface. If automation can deploy production software, create cloud identities, or change trust policies, its credential deserves the same scrutiny as a human administrator account, and probably more. Automation works at three in the morning without getting tired or suspicious.
A firewall became the route to the virtual servers
Check Point disclosed a separate active attack chain against internet-facing Security Gateways. Attackers used two PHP2Shell flaws for initial access, escalated to root through a Gaia OS vulnerability, installed persistence, and in some cases moved by SSH to VMware ESXi hosts and deployed a wiper. Check Point says the destructive activity was limited and affected customers were contacted, but the mechanism is the lesson.
The firewall did not contain the attacker. It became a trusted position from which to reach the virtualization layer. Install the November Jumbo Hotfix and review the indicators of compromise before treating the patch as proof that nothing happened. Then ask who can erase the backups as well as the live virtual machines. Recovery stored behind the same compromised administrative authority is less independent than it looks.
Backups & recovery
Backups are comforting right up until you need one. Raymond Tec helps small businesses build practical backup and recovery plans — including the decidedly unglamorous part where we make sure the thing can actually be restored.
OpenAI began rolling out ChatGPT Wallet to adult Pro subscribers in the United States. A user can add an eligible card, ask ChatGPT to shop, review the item and total, and approve the purchase with a security check such as Face ID or a fingerprint.
The payment mechanism is tokenization. After approval, the card network creates a substitute credential with its own number, security code, and expiration date. The merchant charges the existing card through the token, while the shopping agent and merchant don’t receive the actual card number.
That reduces one risk. It doesn’t decide whether the merchant is reliable, whether the return policy is sane, or whether the agent chose the correct variation. Every purchase currently requires the user’s approval. OpenAI says later versions may allow preapproval within spending limits, categories, or time periods. That will be the more consequential change, because the control moves from “approve this purchase” to “approve a class of future decisions.”
Small ecommerce businesses don’t need to rebuild around a limited rollout. They do need accurate product names, variants, stock, shipping costs, and return terms. The agent can’t rescue a store where “Option 3” means blue, medium, except on Tuesdays. Ask the payment and fraud providers how agent transactions are identified, and watch whether bot protection blocks signed shopping requests that customers intended to send.
Ecommerce development
A good ecommerce site has to work in front of the homepage and behind it. Raymond Tec designs and builds ecommerce sites on Shopify, WooCommerce, Magento / Adobe Commerce, BigCommerce, and other major platforms — including the storefront, product catalogs, inventory, navigation, integrations, fulfillment workflows, and the machinery customers hopefully never have to think about.
A Privacy Toggle Has to Change What the Website Does
California’s privacy regulator ordered FedEx to pay nearly $3 million after finding that its website privacy controls didn’t consistently match the website’s behavior. The decision says device or user identifiers were shared with advertising platforms, some advertising data continued after visitors disabled cookie categories, and the site failed to honor Global Privacy Control signals.
Global Privacy Control is a browser or extension signal telling websites the user wants to opt out of sale or sharing. It isn’t a magical universal privacy switch, but covered businesses have to recognize it. More broadly, a consent banner is not complete because the toggle changes color. Every relevant tag, pixel, and server-side connection has to change behavior too.
Smaller businesses may fall under different legal thresholds, but the engineering lesson does not depend on company size. Keep an inventory of the analytics and advertising tools that run, test the choices a visitor sees, and retest after plugins, tag managers, and marketing integrations change. A banner that says “off” while the same data keeps leaving is theater with rounded corners.
Pennsylvania businesses also have a December 21 deadline approaching under the Artificial Intelligence Deceptive Advertising Act. The law targets AI-generated or substantially modified advertising that creates a false impression about the nature, origin, or authenticity of a product or service. The useful internal question isn’t “Did AI touch the file?” It is “Did the finished ad change what a reasonable customer would believe they are buying?” Add that question to the approval checklist and keep the original assets.
The rules around technology matter too
Platforms, privacy, speech, competition, surveillance, copyright, and regulation increasingly determine what technology companies can build and what the rest of us have to live with. Browse more Raymond Tec News for practical coverage of technology policy and digital rights.
Something Good: A Cell-by-Cell Map of the Brain’s Behavioral Gatekeeper
The NIH BRAIN Initiative published a 27-paper collection mapping the basal ganglia in unprecedented detail. The basal ganglia are clusters of brain structures involved in turning intentions into actions, learning, focus, and movement. Researchers used single-cell sequencing on donated human brain tissue, then linked cell identities and activity to three-dimensional locations. The resulting atlases cover the mature brain, early development, and comparable regions in mice, marmosets, and other nonhuman primates.
This is not a new treatment. It is a reference map, and that distinction makes the work more interesting, not less. Researchers have already used the atlas to identify changes in cell populations deeper in the brains of people with autism that standard approaches may have missed. In mouse models of Parkinson’s disease, another team used a genetic switch to target specific striatum cells and boost a dopamine precursor while reducing common side effects.
The data, preprocessing software, and a web-based exploration tool are publicly available. That gives researchers a shared coordinate system for asking much more precise questions about Parkinson’s, Huntington’s, autism, addiction, and other disorders. Science often advances because somebody builds the map everyone else can use. This is one of those maps.
What to Watch Next Week
Watch for additional Fireworks customer disclosures, because Notion, Uber, and Cursor may not define the full customer impact. N-able and ConnectWise are still investigating, so remote-management operators should follow current advisories instead of relying on the last instruction they remember. We also need a fuller FAA account of why one router change propagated nationally and what the modernization plan changes about that failure mode.
The larger lesson is already clear enough to use. Inventory the identities that can act across systems, reduce what each one can reach, and make sure recovery and investigation do not depend entirely on the same trusted path. Trust is necessary. Unlimited, permanent, poorly observed trust is a design choice.
Still in a reading mood? The Raymond Tec News archive covers security, AI, small-business technology, policy, and the places technology collides with ordinary life — without requiring a computer-science degree to get through it.
Sources / Further Reading
Fireworks AI and third-party access
- Raymond Tec Daily Brief: Notion’s Fireworks AI exposure
- Raymond Tec Daily Brief: Uber’s Fireworks AI exposure
- Raymond Tec Daily Brief: Cursor’s Fireworks AI attack chain
- Fireworks AI: November 9 incident update
- Notion: Fireworks AI incident advisory
- Cursor: investigation and attack-chain update
- Cursor: affected extension versions and remediation
- TechCrunch: Uber driver-license exposure
N-central and ScreenConnect
- Raymond Tec Daily Brief: the N-central shutdown
- Raymond Tec Daily Brief: N-central attacks after HF2
- Huntress: Able2Shell, HF2, and Legacy API guidance
- N-able: current N-central incident updates
- ConnectWise: ScreenConnect investigation and required actions
- Huntress: SCKit campaign updates and indicators
Outages, developer access, and recovery
- Raymond Tec Daily Brief: FAA, Check Point, Microsoft 365, and PostgreSQL
- U.S. DOT: November 14 FAA restoration update
- Aikido Security: wfx-loot technical report
- Check Point: active exploitation advisory
- Check Point Research: intrusion and wiper analysis
Commerce, privacy, and advertising
- Raymond Tec Daily Brief: FedEx privacy controls
- OpenAI: Introducing ChatGPT Wallet
- OpenAI: ChatGPT Wallet Privacy Notice
- California Privacy Protection Agency: FedEx enforcement announcement
- Pennsylvania General Assembly: Artificial Intelligence Deceptive Advertising Act
Something Good
Photo by Christina Morillo on Pexels.
