๐ต๏ธ RESEARCH & DEEP DIVES
Zero-day in Meta's Muse macOS AI agent lets local malware hijack the assistant and steal credentials
A zero-day in Meta's Muse AI agent for macOS lets unprivileged local malware redirect the agent's dictation endpoint to hijack it, inject prompts, and steal authentication data.- Affects Meta's Muse AI agent for macOS, which holds broad delegated permissions over files, email, calendars, browser, purchases, and linked devices.
- The undocumented endo_voyager_dictation_endpoint config setting can be modified by an unprivileged local process, redirecting Muse's dictation traffic to an attacker-controlled server.
- Attack flow: malware already running as the local user (no RCE on a clean Mac) changes the endpoint, then captures dictated audio/prompts, injects malicious instructions, and obtains account authentication data.
- Disclosed by Patrick Wardle (Objective-See) with a proof-of-concept named 'not-a-mused' implementing a subset of Muse's 50+ commands; demos showed querying linked devices' location and triggering BLE scans.
- Meta had not publicly responded; its bug bounty offers up to $300,000 for qualifying Muse flaws.
Sources ๐ Source: github.com ยท ๐ Coverage: cybersecuritynews.com ยท ๐ via Cyber Security News, cryptika.com (discovered)Study: 5.65% of enterprise web apps serve live credentials in production JavaScript bundles
An IEEE Access study documents runtime credential exposure in production web apps that static secret scanners miss.- Applies to organizations running Azure-heavy enterprise web applications that ship JavaScript bundles to browsers.
- Azure AD client credentials and APIM subscription keys were found live in production JavaScript bundles, enabling account takeover and mass data exposure.
- An authorized engagement across ~2,000 enterprise web assets found 113 (5.65%) serving live credentials; 73.3% of exposed apps had the full Azure AD token-mint chain co-located in one browser-reachable bundle.
- CryptoJS-encrypted configuration defeats all static scanners since credentials exist only after runtime decryption with a co-located key; 13.9% of the GT-194 ground truth was found only by manual analysis.
- Best static scanner recovered 36.6% of secrets and best runtime-aware scanner 77.8% (F1 = 0.818); combined coverage plateaus at 86.1% across nine evaluated scanners.
SoK paper systematizes the Linux kernel bug lifecycle from discovery to patch deployment
Researchers have published a SoK systematizing the Linux kernel bug lifecycle and identifying why the crash-to-patch gap persists.- Applies to the Linux kernel and automated bug-finding systems such as syzbot, which now surface bugs faster than downstream processes can absorb.
- Covers five stages of the bug lifecycle: discovery, triage, patch generation, patch validation, and integration.
- Later stages remain far less automated, so fixed bugs often stay open for weeks, need review-driven revisions, or lack reproducers.
- Kernel-specific challenges include concurrency, implicit invariants, cross-syscall state, hardware dependence, lack of fault isolation, and architecture/configuration multiplicity.
- Authors argue repair and validation tools wrongly assume reproducers, localized root causes, and correctness oracles that many real bug reports lack.
Inside BambooToken's Linux implant: shell and file control over MQTT
Researchers reverse-engineered the BambooToken Linux backdoor, which uses MQTT with XOR-obfuscated messages for shell and file control.- Applies to Linux x86-64 hosts; the analyzed sample is a stripped, statically linked C++ ELF (SHA-256 bf0681e43f51e98fe7ec24bf73a43facc66b0feb325a3a835f8f01b2861b926e).
- The implant connects to embedded MQTT broker live-hk[.]c2iznja[.]com on TCP port 2883 and subscribes to a fixed group topic {534E19D5-434B-4cad-A0C2-8D75E0B2FBFC}.
- It resolves into three components: a host inventory collector, a shell command worker using /bin/sh -c via popen, and a file manager supporting dir, download, upload, and delete operations.
- Messages are XORed with a repeating 37-byte key; topics are XORed per segment and hex-encoded, and the sample supports background daemonization via a double-fork.
- The sample's SHA-256 is listed in Black Lotus Labs' published BambooToken IOC collection, confirming family attribution.
- Decoded file-listing responses use the misspelled field 'flies', a useful traffic discriminator.