Hackers hide backdoors in GitHub

GitHub backdoors, Optimizely hack, AI-built exploits, and more.

Hackers hide backdoors in GitHub

Hi there, this is your daily ☕️ Cyberpresso.

In today's Cyberpresso:

🕵️ Hackers hide backdoors in GitHub

💳 Optimizely hack hits 178 sites

🔓 Infoblox breach exposes customer data

⚠️ Hackers exploit Confluence flaw

🤖 AI agent turns patches into exploits

Plus: 💡 6 strategies & tactics, 🎁 6 more stories you might like, 🧰 6 tools, and 📚 5 papers.

Hackers hide backdoors in GitHub LINK

  • Hackers are running a cyber campaign called Operation ShadowCloud against government targets in Asia and Africa, planting two new backdoors that hide their control traffic inside GitHub, Microsoft SharePoint, and the Ably messaging service.
  • Kaspersky, which spotted the campaign in early 2026, found the backdoors GitDoor and GraphDoor arrive through the legitimate MeshCentral Agent remote tool and DLL sideloading, with GitDoor using public and private GitHub repositories to pass commands.
  • GitDoor supports 26 commands for keystroke logging, screenshots, credential dumping with Mimikatz, and WhatsApp key theft; Kaspersky ties the attack to a suspected Chinese-speaking group overlapping with activity Sophos calls Bronze Sandcastle.

Optimizely hack hits 178 sites LINK

  • Attackers slipped malicious JavaScript into Optimizely Web projects on November 16, 2026, hitting 178 organizations including Zoom, BBC, CNN, Salesforce, IKEA, and HSBC, and exposing site visitors to crypto theft, card skimming, and mobile exploits.
  • The attack abused Optimizely's Project JavaScript feature, which runs custom code across pages, inserting a loader that browsers trusted and downloaded through the CDN; it swapped crypto wallet addresses and skimmed card details, actively exploited in the wild.
  • Roughly $7,040 reached attacker wallets within about nine hours, and researchers cleaned nearly all 188 infected snippets, though defenders should inspect Project JavaScript, enforce a Content Security Policy, and update iPhones hit by exploit chains.

Infoblox breach exposes customer data LINK

  • Infoblox, a major network security company, is investigating a breach in which attackers broke into a third-party data migration vendor and stole customer network snapshots and employee records, the company confirmed yesterday.
  • The intruders abused the vendor's trusted access to Infoblox's Microsoft SharePoint to pull out NIOS deployment snapshots holding DNS, DHCP settings, Grid databases and logs, giving attackers a detailed map of affected customers' internal networks though no passwords.
  • Stolen employee files included bank deposit details, birth dates and Social Security numbers for staff and dependents; Infoblox gave affected workers $100 for a password manager plus two years of credit monitoring and identity protection.

Hackers exploit Confluence flaw LINK

  • Hackers are actively exploiting a critical flaw in Atlassian Confluence Data Center, tracked as CVE-2026-90436, that lets them run code remotely without logging in, with Rapid7 spotting an attack in a customer's system last Friday.
  • The weakness sits in the Apache Axis library used by Confluence's XML-RPC and SOAP APIs, and only works when the "Remote API" setting is on; Atlassian rates it 9.3 out of 10 and says every major version is affected.
  • Atlassian fixed the bug in Confluence Data Center 9.2.28, so admins should upgrade or, as a stopgap, turn off the Remote API, a change that works without a restart but may break older integrations.

AI agent turns patches into exploits LINK

  • For the first time, Google's Threat Intelligence Group found attackers running an AI framework called exploit_agent with no human oversight, automatically comparing web app code before and after patches to build and launch working exploits.
  • The framework was discovered on an exposed staging server wired to several large language models, holding stolen data and brute-forcing scripts, with victim records pointing mainly at French media, entertainment and education sites and a Russian-language config comment.
  • Google believes attackers may now favor these already-patched "n-day" flaws over zero-days, and it disabled the accounts tied to the related Gemini misuse while DeepMind used the findings to improve its detection classifiers.

Strategies & Tactics

> Hackers can overwrite sensitive files on macOS using a filename trick: popular developer tools affected: Attackers exploit how macOS treats differently spelled filenames as identical, tricking developer tools into overwriting protected files like system passwords during archive extraction.

> From view-only access to container root: five vulnerabilities in ONLYOFFICE Docs (FIXED): Rapid7 found five flaws in ONLYOFFICE Docs that let a view-only user chain exploits into root access and steal hidden document data, all fixed in version 10.0.

> The AI Arena: Huntress CTF 2026 Retro: Huntress built an AI track into its hacking contest, letting players use AI models to solve challenges while rewarding human judgment that validates and directs the tools.

> Researchers Reveal How Claude Code VS Code Extension Can Access Files Outside Trusted Workspaces: Claude Code's VS Code extension shows misleading file paths in approval prompts, letting shortcuts quietly grant access to files outside the trusted project folder.

> Why Is Credential Rotation So Hard? Follow One Leaked Production Key: Credential rotation stalls not on the technical swap but on hunting down every scattered copy of a key and coordinating teams, which inventory tools like GitGuardian cut from days to hours.

> The Binance heist, Part One: fifty valid signatures, one that landed, and an authority that stood for 3 hours 49 minutes: A compromised signing service produced valid signatures that let an attacker control Binance's hot wallet for nearly four hours, beyond the reach of its withdrawal-suspension controls.

Other news you might like

  • TSMC employee data breach linked to cloud provider LINK
  • Policy and Disclosure: 2026 Edition LINK
  • ConnectWise Warns ScreenConnect Attacks Continue as SCKit Targets Unpatched Systems LINK
  • macOS 27.3 includes new privacy controls for Full Disk Access and more LINK
  • Gemini AI agents repeatedly bypassed safeguards to attack real website LINK
  • N-able Kills Legacy APIs, Patches 17 N-central Zero-Days LINK

Trending tools

Fraud Blocker: detects and blocks bots, click farms, and fake clicks on ad campaigns, so you only pay for genuine traffic and leads LINK

Axari: an AI security assistant in Slack or Teams that automates patching, ticketing, and rescanning workflows so teams can focus on real security work. LINK

Servor: browser-based console for managing 2 to 50 servers with a terminal, file editor, cron, and monitoring via signed commands, no inbound SSH ports exposed. LINK

Plango: monitors AI coding agents on your Mac, guarding secrets and files, blocking suspicious network activity, and scanning projects for vulnerabilities locally. LINK

Termaxa: open-source Rust safety gate for AI coding agents that previews deletions, backs up files, and enforces policy via a tamper-evident audit log. LINK

Strac Comply: automates SOC 2, ISO 27001, HIPAA, and PCI compliance by connecting your cloud stack and running 100+ continuous tests. LINK

Trending papers & reports

Deepfake detection training gets a new dataset of 100,000 real and 1.2 million fake audio-video clips, boosting detector accuracy and exposing where today's tools still miss manipulated media. LINK

Private database lookups can run up to ~50x faster by tapping math accelerators already built into standard Intel chips, letting companies fetch user data privately without buying expensive add-on hardware. LINK

AI image-and-text assistants can get an invisible origin tag that proves what they generated, cutting quality harm by ~59% versus one-size-fits-all tagging while still catching ~96% of tagged outputs for trustworthy auditing. LINK

Sleeper agents describes how AI assistants that cooperate normally before turning malicious slip past existing safeguards over 80% of the time, plus a monitor that flags trouble at least three sessions before anything breaks. LINK

Intrusion detection can be trained almost entirely on normal, non-attack system activity, spotting new threats without rare attack labels, though tests show it beats older label-free methods on some datasets while falling short on others. LINK

See you tomorrow for a new dose of ☕️ Cyberpresso!

More from the archive