TRENDING
A black manual telephone switchboard with rows of orange jacks and red patch cords plugged into the lower row, above a rotary dial and a handset
November 16, 2026
How to Route Alerts in Alertmanager So Pages Reach the On-Call and Tickets Reach the Queue, and Test Grouping, Inhibition, and Silences Before an Outage Does
A brass nail with a threaded screw section lying on graph paper, a fastener that is both nail and screw
November 16, 2026
Ai2’s XYBench Turns the XY Problem Into a Benchmark, but the Model That Drafted Every Rubric Also Judges the Answers, the Human Baseline Is Graded on Rubrics Read Off Its Own Answers, and the Emphasis Judge Disagrees With the Annotators on 20 to 26 Percent of Validation Responses
An amber warning beacon on a white surface, a flashing-light stand-in for the FBI’s FLASH alert
November 16, 2026
The FBI’s ARTEX Warning Says a Supply Chain Compromise “Enabled Access” to 2,250 Tenants Where Microsoft Counted Sign-In Attempts and Cites “Over 1,600” Consoles Where Censys Counted 1,465, and a Microsoft Post It Cites Mistypes One Indicator in Its Hunting Query
Two eggs and a colour-changing egg timer in a saucepan of water on an induction hob
November 16, 2026
How to Set gRPC Deadlines and Retries in Python, Pass Them Down a Call Chain, and Stop Retries From Charging a Card Twice
A whole lemon and a halved lemon side by side on a black background
November 16, 2026
IBM’s LLMON Turns Prompt-Injection Defense Into a Token Layout, but the “Nearly Zero” Chart Is One Benchmark, the Paper’s Own Red Team Lets 15% to 20% of Indirect Attacks Through, and the Shipped Template Lets a Document Type the Delimiters
16 Nov 2026
SXZ.io SXZ.io
  • Home
Search the Site
Popular Searches:
Technology Amazon AI
Recent Posts
A manual paper cutter with a wooden grid board, a clear plastic guard and a long cutting arm, a picture of text being trimmed to fit
How to Stop Ollama From Silently Cutting Your Prompt, Find the Real Context Window, and Set num_ctx Without Running Out of Memory
November 16, 2026
A brown llama and her gray cria walking across a salt flat beside a red lake in Bolivia
llama.cpp’s v1.0.0 Promises Major Bumps for API and ABI Breaks, but Its Rule Table Lists Only the C API, libmtmd Shares the Version While Its Header Warns of Breaking Changes, and the Release Page Carries a 7-Byte Pointer to a Nightly Build Instead of Binaries
November 16, 2026
A stack of chocolate chip cookies on a white plate beside a cup of tea
NoisyTCP Can Crash a Linux Server Over the Network, but the Fix Has Reached Only Three of the Seven Kernel Lines Still Being Released, Red Hat and Ubuntu Have No Update, and the Researcher’s Own Timings Run From 5 Seconds to 14 Hours
November 16, 2026
SXZ.io SXZ.io
  • Home

Categories

Articles 302 Posts
News 307 Posts
Learning Hub 274 Posts
Home/News/Cursor Traces the Tampered Claude Code Extension to a Key Exposed at Fireworks AI and Agent Traces Holding Unredacted Credentials, but Neither Company Says When the Key Leaked, and Fireworks’ Bulletins Never Name Cursor
News

Cursor Traces the Tampered Claude Code Extension to a Key Exposed at Fireworks AI and Agent Traces Holding Unredacted Credentials, but Neither Company Says When the Key Leaked, and Fireworks’ Bulletins Never Name Cursor

Cursor’s November 13 report says the tampered Claude Code extension began with a storage key exposed at Fireworks AI and agent traces holding unredacted credentials; neither company says when the key...

November 14, 2026 10 Min Read
26

Cursor has published its account of how a tampered copy of Anthropic’s Claude Code extension reached its VS Code extension marketplace on October 29. The report, dated November 13, says the chain began at Fireworks AI, the vendor that handled some of its model training and inference. An attacker obtained an access key Cursor had issued to Fireworks, used it to copy coding-agent traces from a read-only storage bucket, found credentials in those traces that had not been redacted, reached some of Cursor’s SaaS tools with them, and then used a stolen credential to place the tampered copies directly on the cache that serves the marketplace. Cursor says at most 5,000 users potentially received the tampered build and that the last attacker activity it saw was on November 1. BleepingComputer reported the update on November 14.

Table Of Content

  • What Cursor says happened, and what each step leaves out
  • What Fireworks has published
  • The dates side by side
  • Which of our October 30 questions the report answers
  • What we checked ourselves
  • Where the accounts differ
  • What to do now
  • Sources and method

The report answers the question our October 30 post left open, which was how the actor reached the marketplace. It leaves most of the other questions open, and it gives no date for any step before October 29. Fireworks’ two public security bulletins never mention Cursor, a storage bucket or an extension, and neither company says when the key was exposed. We also downloaded and hashed the genuine 2.1.312 files from Open VSX: they match Open VSX’s own checksums and differ from the “Modified VSIX” hashes in Cursor’s notice, and Open VSX still has no 2.1.313.

What Cursor says happened, and what each step leaves out

The table follows the report’s own order. The middle column quotes Cursor’s report. The right column is what the report does not say, which is our reading.

Step What Cursor says What the report leaves out
1. Key exposed “An attacker obtained an access key we had issued to Fireworks. The key was exposed by Fireworks” No date, no named Fireworks incident, and no account of how Fireworks held the key.
2. Bucket copied The key opened “a read-only storage bucket we used to share data with them” that held “limited coding agent traces, mostly from 2025, from sessions not in Privacy Mode, which we were permitted to use for model training.” No date for the copy, no trace count, no word on how long the key stayed valid, and no date for when Cursor’s monitoring flagged the activity.
3. Credentials found “A small number of those traces contained credentials that had not been redacted; less than 0.01% of customers had API credentials exposed.” No absolute count and no total to apply the percentage to, and no list of which services issued the credentials.
4. SaaS tools reached The attacker “reached select SaaS tools used by Cursor” and took “select archived non-production Cursor source code repositories (some containing business contact information), limited customer support data and email conversations, product usage metrics and contact details, summaries of some coding agent conversations, and some credentials users had sent to the Cursor support team.” Which tools, and whether any of them held the credential used on October 29.
5. Cache written On October 29, “using a stolen credential from the exposed data from our internal system, the attacker placed malicious, unauthorized copies of Anthropic’s Claude Code extension directly on the cache used to serve the Cursor VS Code extension marketplace.” Which credential, whether it came out of the traces or out of a SaaS tool reached later, and how a write to the cache produced a modified build under the number of a genuine release.
6. Reach “At most 5,000 users who requested specific cached versions of the extension on October 29 potentially received the malicious build.” How many ran it, and whether anyone logged in over the tunnel.
7. Containment “The last attacker activity we have seen was on November 1, and it was blocked.” What that activity was.

Cursor adds that Anthropic “did not publish this copy, and this was not a vulnerability in Anthropic’s product.” It says it ships its app “from a separate code hosting system that the attacker never accessed” and that “no release signing keys were exposed.” For people who were not in Privacy Mode, it says traces were “exposed for fewer than 0.05% of users,” that it has notified each of them, and that “If you did not receive a notification, your data was not in these traces.” That is the vendor’s statement, and we cannot confirm it from outside.

What Fireworks has published

Fireworks has two bulletins on its incident, an October one (published on October 9, last updated October 10 at 11:16 PDT) and an update dated November 9. A text search of both finds no mention of Cursor, Anthropic, a bucket, an extension or a marketplace.

The October bulletin says that “On October 6, Fireworks identified a security incident involving unauthorized use of internal Fireworks credentials.” An unauthorized party “accessed environment variables used in evaluation jobs in a limited number of customer accounts.” Fireworks said it had “not detected evidence of unauthorized access to customer models, datasets, training data, personal information, or inference traffic within the Fireworks security perimeter or control.”

The November update widens that. It says “only Fireworks stored secrets (e.g. used in evaluation jobs) were known to have been accessed” when the first bulletin went out, and that “For a limited set of accounts, we also discovered impact to secrets, datasets, and incomplete models.” It adds a second incident: “A third-party cloud provider that hosted Fireworks R&D machines experienced a security incident on October 23. On October 24, the unauthorized third party used it to reach Fireworks development machines and then run an unauthorized workload that led to additional customer data impact.” Fireworks says it detected that activity the same day and that “As of October 25, we have not detected further evidence of successful unauthorized access by the threat actor on the Fireworks platform customers use.”

Two readings of ours, both unconfirmed by either company. First, the no-evidence sentence in the October bulletin is limited to data “within the Fireworks security perimeter or control.” A bucket Cursor used to share data, read with a key Cursor issued, arguably sits outside that wording, which may be why neither bulletin mentions it. Second, a key that Cursor issued to Fireworks and that Fireworks stored would fit the update’s phrase “Fireworks stored secrets.” Neither company says either thing.

The dates side by side

Times are UTC unless marked. Pacific time was UTC minus 7 through October 31, and clocks changed on November 1. Weekdays and gaps are our arithmetic.

When What happened Source
Oct 6 (Tue), 16:00 Fireworks identifies unauthorized use of internal credentials and begins containment (9 a.m. PDT). Fireworks, October bulletin
Oct 10 (Sat), 05:00 Fireworks publishes its first bulletin (10 p.m. PDT on Oct 9). Fireworks, October bulletin
Oct 23 (Fri) A third-party cloud provider that hosted Fireworks R&D machines has a security incident. Fireworks, update
Oct 24 (Sat) The unauthorized party reaches Fireworks development machines and runs an unauthorized workload; Fireworks says it detected this the same day. Fireworks, update
Oct 29 (Thu), 09:27 to 20:35 Tampered copies of the Claude Code extension are available in Cursor’s marketplace (2:27 a.m. to 1:35 p.m. PDT), 11 hours and 8 minutes. Cursor, notice
Nov 1 (Sun) Last attacker activity Cursor saw; it was blocked. Cursor, report
Nov 2 (Mon), evening Fireworks tells Notion that a small set of data was accessed (time zone not stated). Notion disconnects from Fireworks by the morning of Nov 3. Notion, advisory
Nov 6 (Fri) and Nov 8 (Sun) Fireworks tells Notion that malicious actors may have accessed potential customer data, and confirms the dataset two days later. Notion, advisory
Nov 9 (Mon) Fireworks publishes its update; Notion publishes its advisory. Fireworks, Notion
Nov 13 (Fri) Cursor publishes its report, 15 days after the notice. Cursor, report

From the day Fireworks identified its first incident (October 6) to the poisoning is 23 days, and from the day the second reached its development machines (October 24) it is 5 days. Cursor does not say which incident exposed the key, when the bucket was read, or whether Fireworks warned it. Cursor says only that “Our monitoring systems identified the unauthorized activity.” Notion’s advisory, by contrast, dates each warning it got from Fireworks. Fireworks was also one of the platforms in the “Black October” list our November 1 post compared, and BleepingComputer places Cursor inside the same wave.

Which of our October 30 questions the report answers

Open question on October 30 After the November 13 report
How the actor reached the marketplace Partly answered: a Fireworks-held key, then a bucket of traces, then credentials, then SaaS tools, then a stolen credential used on the cache. The credential itself is not identified.
How many machines installed a tampered build The 5,000 ceiling is back, now defined as users who requested cached versions on October 29 and “potentially received” the build. No count of machines that ran it.
Whether anyone logged in Not addressed.
Whether Linux builds were served Not addressed. The report names the versions but no platforms.
Which version the rebuilt marketplace serves now Not addressed. The report says Cursor “reviewed and rebuilt our VS Code extension marketplace.”
Why 2.1.313 is in no official channel Not addressed. The report calls 2.1.312 and 2.1.313 tampered versions.

What we checked ourselves

Cursor says the marketplace source, Open VSX, “never had the malware.” We queried Open VSX’s public API on November 14 between roughly 19:05 and 19:25 UTC. This is what it holds for Anthropic’s Claude Code extension.

Version What Open VSX returns
2.1.311 Published Oct 26 at 21:53 UTC.
2.1.312 Published Oct 27 at 20:26 UTC, about 37 hours before Cursor’s window opened. The darwin-arm64 build was uploaded at 20:27:13 and the win32-x64 build at 20:27:56.
2.1.313 HTTP 404 for the version and for both platform builds.
2.1.314 Published Oct 30 at 01:07 UTC, 4 hours and 32 minutes after Cursor’s window closed.

Open VSX’s list also has no 2.1.315, 2.1.320 or 2.1.324, so a number missing from its list is not unusual by itself. Here is the check for the two numbers that matter, as a script that needs only Python’s standard library:

import json, urllib.error, urllib.request

base = "https://open-vsx.org/api/Anthropic/claude-code/win32-x64/"
for version in ("2.1.312", "2.1.313"):
    request = urllib.request.Request(base + version, headers={"User-Agent": "version-check"})
    try:
        info = json.load(urllib.request.urlopen(request))
        print(version, "found, build uploaded", info["timestamp"])
    except urllib.error.HTTPError as error:
        print(version, "HTTP", error.code)
2.1.312 found, build uploaded 2026-10-27T20:27:56.403203Z
2.1.313 HTTP 404

For 2.1.312 we then downloaded both genuine builds from the files.download address in each JSON reply, hashed them, and compared the result with the checksum file at the files.sha256 address and with the “Modified VSIX” hashes in Cursor’s notice. The files are 127,409,262 bytes (darwin-arm64) and 141,086,900 bytes (win32-x64).

Open VSX 2.1.312, SHA-256 of the file we downloaded (equals Open VSX's checksum file)
darwin-arm64  b851e43b35b634987ef794fbddc759a7eeabd662088511a276f6063a598373d2
win32-x64     4393b5a468057ec65061083604f2db1f03bdd2725f2a72c485eaae6cffe75ddd

Cursor's notice, "Modified VSIX SHA-256" for 2.1.312
darwin-arm64  fcbb327e38e61c4fec795e0d0591f52cd1014b813d20a5912104953c3a17c364
win32-x64     3de1722901165769893bef57c72594e9a54459faa23f3091b136b79966edd081

No Open VSX hash equals either of Cursor’s modified hashes. That fits Cursor’s statement that Open VSX never held the modified files. It does not show what Cursor’s own cache held before the incident or holds now, because we did not download anything from Cursor’s marketplace. Cursor’s wording is also consistent with the attacker overwriting a cached copy of the genuine 2.1.312 and adding a 2.1.313 that Open VSX does not list, but the report does not say that, so treat it as our inference. Our October 30 post explains why the version number alone cannot clear a 2.1.312 install and how to check an extension folder by content.

Two links are worth knowing about. Cursor’s “exposed by Fireworks” link goes to Fireworks’ bulletin index, not to a bulletin that describes a Cursor key. BleepingComputer’s link to Cursor’s remediation guidance returns HTTP 404 today; the notice is at https://cursor.com/blog/claude-code-extension-security-notice.

Where the accounts differ

  • Whose credentials were stolen. BleepingComputer’s lead says the attackers acted “after stealing credentials from AI infrastructure provider Fireworks AI.” In Cursor’s account the key that came through Fireworks opened a bucket Cursor used to share data, and the credentials that reached Cursor’s SaaS tools came out of the traces in that bucket.
  • Which percentage applies to what. Cursor gives 0.05% of users for traces and 0.01% of customers for credentials, with no totals. The traces sentence opens with “For users not using Privacy Mode,” and the report does not say whether the 0.05% is a share of all users or of those users. The 5,000 figure is the only count in the report, and it is a ceiling.
  • When the notice went out. The report says that “On October 30, we published a security notice on our blog and on X with indicators of compromise, remediation steps and scripts to check your system.” The notice is dated October 29 and carries an October 30, 11:30 a.m. Pacific update that revised the indicators and the affected versions, so the two dates fit different versions of one page. The Internet Archive holds one capture of the report (November 14, 03:12 UTC), so we could not check it for later edits.
  • Who raised the alarm. Cursor says its own monitoring found the activity. Notion says Fireworks warned it on November 2. The report does not say whether Fireworks warned Cursor.

What to do now

  1. If you installed or updated the extension on October 29. Follow Cursor’s notice for versions 2.1.312 and 2.1.313 from its marketplace between 09:27 and 20:35 UTC: look for the svc-telemetry key, run the remediation script for your system, and remember that “Uninstalling the extension alone may not be enough.”
  2. Check your mail from Cursor. The report says it has contacted each affected customer directly, and for the traces it says that if you received no notification, your data was not in them. We cannot confirm that from outside, so do the next step as well.
  3. Rotate what you sent to support. Our advice: any credential you ever emailed or pasted to Cursor’s support team, because the report lists such credentials among the stolen data, and any key you typed into an agent session with Privacy Mode off, especially in 2025, if you cannot rule out that it sat in the shared traces.
  4. If you hand logs or traces to a vendor. Our advice: redact secrets before the data leaves, bind the sharing key to the vendor’s network ranges (Cursor says it has “Expanded network access restrictions across our service providers, where possible, so compromised credentials can’t be used outside their intended location”), and treat a partner’s incident bulletin as a reason to rotate every key you gave that partner, even when the bulletin does not name you. Fireworks’ first bulletin was out 19 days before the poisoning.

Sources and method

We read Cursor’s November 13 report and its October 29 notice, Fireworks’ October bulletin and November update and its bulletin index, Notion’s advisory, and BleepingComputer’s stories on Cursor, Fireworks and Notion. First-hand, we queried Open VSX’s version list and its per-build records, downloaded the two genuine 2.1.312 files and hashed them with SHA-256, searched the saved text of both Fireworks bulletins, and asked the Internet Archive’s index for captures of the report. Weekdays, the 11 hours and 8 minutes, the day counts and the hour gaps are our arithmetic. We did not download anything from Cursor’s marketplace, run any payload or remediation script, or contact any address in the indicator list, and we did not test the Cursor, Fireworks or Notion systems, so every statement about what happened inside them rests on their own accounts.

Tags:

AI Coding AgentsClaude CodeCredential TheftCursorFireworks AISupply Chain SecurityThird-Party Risk

Share

A moss-covered wooden buffer stop with rusty end plates at the end of a disused railway line, with a birch tree growing through the undergrowth beside it
Previous Post

How to Build a Buffer Pool in Python and Stop One Table Scan From Flushing Your Hot Pages With Clock Sweep, a Scan Ring, and Midpoint Insertion

Three buckets on a wet wooden floor collecting water that leaked into a ship interior
Next Post

WIRED’s “AI-Fueled Ransomware Boom” Rests on Leak-Site Counts That Reproduce at 25% to 39%, but the ReliaQuest Report Behind Its New-Group Numbers Calls the Rise Unlikely to Reflect a Real Increase in Activity and Never Mentions AI, and in a Public Tracker 97% of the Rise Comes From Groups First Seen in Q3

No Comment! Be the first one.

Leave a Reply Cancel reply

Your email address will not be published. Required fields are marked *

Latest
16 Nov
How to Stop Ollama From Silently Cutting Your Prompt, Find the Real Context Window, and Set num_ctx Without Running Out of Memory
16 Nov
llama.cpp’s v1.0.0 Promises Major Bumps for API and ABI Breaks, but Its Rule Table Lists Only the C API, libmtmd Shares the Version While Its Header Warns of Breaking Changes, and the Release Page Carries a 7-Byte Pointer to a Nightly Build Instead of Binaries
Trending
November 16, 2026
How to Stop Ollama From Silently Cutting Your Prompt, Find the Real Context Window, and Set num_ctx Without Running Out of Memory
November 16, 2026
llama.cpp’s v1.0.0 Promises Major Bumps for API and ABI Breaks, but Its Rule Table Lists Only the C API, libmtmd Shares the Version While Its Header Warns of Breaking Changes, and the Release Page Carries a 7-Byte Pointer to a Nightly Build Instead of Binaries
November 16, 2026
NoisyTCP Can Crash a Linux Server Over the Network, but the Fix Has Reached Only Three of the Seven Kernel Lines Still Being Released, Red Hat and Ubuntu Have No Update, and the Researcher’s Own Timings Run From 5 Seconds to 14 Hours
November 15, 2026
How to Route Alerts in Alertmanager So Pages Reach the On-Call and Tickets Reach the Queue, and Test Grouping, Inhibition, and Silences Before an Outage Does
November 15, 2026
Ai2’s XYBench Turns the XY Problem Into a Benchmark, but the Model That Drafted Every Rubric Also Judges the Answers, the Human Baseline Is Graded on Rubrics Read Off Its Own Answers, and the Emphasis Judge Disagrees With the Annotators on 20 to 26 Percent of Validation Responses
November 15, 2026
The FBI’s ARTEX Warning Says a Supply Chain Compromise “Enabled Access” to 2,250 Tenants Where Microsoft Counted Sign-In Attempts and Cites “Over 1,600” Consoles Where Censys Counted 1,465, and a Microsoft Post It Cites Mistypes One Indicator in Its Hunting Query

Related Posts

Rows of server racks in a data center representing network infrastructure targeted by botnets
News

C0XMO Botnet Shows Why Old Router Firmware Still Matters

June 7, 2026
Close-up of a USB flash drive, representing physical data-theft risk in office security incidents
News

Fake IT Support Is Now Walking Through the Front Door

June 7, 2026
A phone security app on a smartphone resting on a laptop keyboard.
News

Everest Forms Pro Flaw Is Being Exploited to Create Rogue WordPress Admins

June 7, 2026
A phone secured by a padlock, illustrating AI data-leak containment and security controls.
News

OpenAI’s Lockdown Mode Is a Data-Leak Brake, Not a Prompt-Injection Cure

June 8, 2026
SXZ.io SXZ.io
  • [email protected]

Categories

Articles
Learning Hub
News

All Rights Reserved by SXZ.io ©2026