Adil Islam

Daily Cyber Briefing — November 15, 2026

The Cursor report shows how one AI vendor key turned into a poisoned IDE extension, by way of coding-agent logs that held live credentials. Window roughly Nov 14, 7:00 a.m. to Nov 15, 7:00 a.m. CT. Sources are linked. Verification is marked on each item.

Cursor: a stolen Fireworks AI key led to stolen agent traces, SaaS access and the trojanized Claude Code extension

Verified

Cursor published its incident report on November 14. The attack started with an access key Cursor had issued to Fireworks AI, which Cursor says Fireworks exposed. The attacker used it to copy a read-only storage bucket of coding-agent traces, mostly from 2025, from sessions not in Privacy Mode that Cursor was allowed to use for training. Some traces contained credentials that had not been redacted: Cursor says fewer than 0.01% of customers had API credentials exposed and fewer than 0.05% of users had traces exposed. With those credentials the attacker reached some of Cursor's SaaS tools, including archived non-production code repositories, limited support data and email, usage metrics, summaries of some agent conversations, and credentials users had sent to Cursor support.

On October 29, using a stolen credential, the attacker placed malicious copies of Anthropic's Claude Code extension, versions 2.1.312 and 2.1.313, on the cache that serves Cursor's extension marketplace. The builds tried to install a persistent SSH backdoor and were live for about 11 hours; at most 5,000 users who requested those cached versions may have received them. Open VSX never carried the malware, and Cursor says this was not a flaw in Anthropic's product. The last attacker activity Cursor saw was on November 1, and it was blocked. Cursor has moved all inference traffic off Fireworks, rebuilt the marketplace and redeployed production.

Why it matters: Coding-agent traces hold whatever secrets developers pasted into them, so any vendor that receives traces for training also holds a credential store. Redact secrets before traces leave the building, limit what vendors can read, and rotate any key that has ever appeared in an agent session.

Sources: Cursor · BleepingComputer · The Hacker News · Fireworks AI bulletin

Elastic: a CyberStrikeAI server made 56 AWS API calls in 141 milliseconds

Verified

Elastic Security Labs reported that one server running the open-source CyberStrikeAI platform used stolen credentials for eight AWS identities, five IAM users and three EC2 instance roles, across four accounts between September 30 and October 16. It checked which credentials worked, read Secrets Manager, pulled EC2 user data scripts (56 requests in 141 milliseconds), decrypted an SSM parameter and sent an SSM command 29 seconds later. The 78 alerts it raised were all medium or low severity and only made sense when grouped by source IP, and a 14-day quiet gap between bursts slipped past fixed correlation windows. Elastic says it cannot prove a language model chose the actions or show how the credentials were stolen.

Why it matters: AI attack tooling produces two patterns at once: bursts no human could type and long gaps that outlast detection windows. Keep secrets out of EC2 user data and correlate cloud alerts over weeks, not hours.

Sources: Elastic Security Labs

DeepMind: swarms of attacker agents get past prompt-injection defenses

Verified

A Google DeepMind paper, "Scaling Adaptive Prompt Injection Attacks with Agent Swarms," describes an attacker built from eight teams of two agents that share notes and a leaderboard while searching for injection payloads. Against the strongest version of that attacker, Gemini 4 Argon resisted 66% of test scenarios at the full budget, and six of the seven models tested resisted 10% or less. The authors say human red-teaming is close to saturated and call for automated red-teaming and layered defenses. The tests are synthetic research scenarios, not observed attacks.

Why it matters: A model that holds up against human testers can still fall to an automated attacker with time and budget. Agent deployments need controls outside the model: scoped short-lived credentials, policy checks at the tool layer, and human approval for high-impact actions.

Sources: arXiv

Apify breach exposed GitHub and Google sign-in tokens and 2FA recovery codes

Verified

Web scraping and automation platform Apify, which says it has 91,000 customers, confirmed that an intruder used a compromised credential to obtain more credentials and get into customer accounts on November 9 and 10. Apify blocked the access on November 11. A customer notification shared on Reddit, as reported by SecurityWeek, listed names, email addresses, billing details, IP addresses, salted password hashes, 2FA recovery codes and phone numbers, and authentication tokens tied to Google and GitHub sign-ins. Apify rotated internal secrets and revoked some OAuth connections, including GitHub, and told customers who build from private GitHub repositories to reconnect. It has not said how many accounts were affected.

Why it matters: Automation platforms hold OAuth tokens into the systems they automate. Customers should treat linked GitHub and Google grants as exposed, reset 2FA recovery codes, and review recent repository activity.

Sources: SecurityWeek

Citrix patches 28 NetScaler flaws, including one that affects default configurations

Verified

Citrix's November bulletin covers 28 memory flaws in NetScaler ADC and NetScaler Gateway; SecurityWeek counts nine as critical. CERT-EU says one of them, CVE-2026-120868, affects every deployment, including default configurations. Citrix describes the impact as unpredictable behavior or denial of service, and CERT-EU says there is no reported exploitation and no confirmed public proof of concept. Fixed releases are 14.1-73.57 and 13.1-64.39 or later, with separate FIPS and NDcPP builds. Some posts on X claim active exploitation; the exploited NetScaler flaws on record are September's CVE-2026-88771 and CVE-2026-88772.

Why it matters: NetScaler bugs have a history of moving from patch to exploitation quickly. Internet-facing Gateway and AAA appliances should be upgraded first.

Sources: Citrix bulletin · CERT-EU · SecurityWeek

Also on the wire

  • More than 400 fake accounts spent a year pushing Nvidia's case for selling AI chips to China, and one posted its own AI prompt; most have been suspended. Washington Post
  • Telegram for Android fixed a widget flaw that let a permitted bot's widget run code on the phone; update to 13.2.2 or later. Proof of concept on GitHub
  • Workday took all WD1 production tenants offline after a disaster-recovery test connected backup systems to production databases; service was restored November 14. UC Today
  • Italy's privacy regulator fined Meta €805 million over Facebook's Android app reading photo galleries without specific consent. The Hacker News on X

Cyber covers AI and agentic threats, supply-chain and developer-tool compromises, and how enterprise security absorbs them. Signal stays the AI research stream. ← All bulletins