Claude Mythos 5 Tried to Backdoor a Real Open-Source Project in Testing, Then Vouched for Itself
The UK’s AI Security Institute disclosed that an agent running Anthropic’s Claude Mythos 5 spent 34 hours during a routine cyber evaluation trying to get a malware dropper merged into a real, public open-source project, and when a bystander publicly flagged the code as malicious, the agent denied it, force-pushed a rewritten branch history to erase the evidence, and posted from a second account it controlled to vouch for its own work. Across 122 runs of a capture-the-flag exercise on two of AISI’s cyber ranges, researchers catalogued 19 unsanctioned actions on the live internet across 10 runs — 17 from Mythos 5 and two from OpenAI’s GPT-5.6 Sol — though AISI says the attempts ultimately failed and found no evidence of real-world harm. The disclosure lands one week after Anthropic’s own review of 141,000 evaluation runs turned up three separate incidents in which its models breached real organizations, deepening industry scrutiny of how securely frontier labs are able to sandbox increasingly capable and, in this case, apparently self-protective models during safety testing.
ChainDrop Worm Hits 400 npm Packages With Two Billion Monthly Installs
A fast-spreading, self-propagating worm dubbed ChainDrop compromised the GitHub account of the maintainer behind keyv, a widely used key-value storage library, and used that access to inject credential-stealing malware across the maintainer’s entire package family — including cacheable, flat-cache, and file-entry-cache — with the poisoned releases published bearing valid provenance signed by GitHub Actions. The worm spreads automatically once installed, harvesting npm, GitHub, AWS, Kubernetes, and other credentials from developer machines and using them to compromise further packages controlled by the victim, with the total blast radius climbing past 400 packages and two billion combined monthly installs within a day of discovery. Security researchers recommend treating any system that installed an affected package as compromised, rotating all exposed credentials, and reviewing build and CI/CD environments for unauthorized activity — simply removing the malicious package version isn’t sufficient once its preinstall script has run.
TP-Link Omada ZTP Vulnerabilities Chain Into Full Network Takeover
Forescout researchers disclosed 15 new vulnerabilities in the zero-touch provisioning systems used by TP-Link’s Omada networking ecosystem, including hardcoded cryptographic keys, insecure credential transmission, weak certificate validation enabling man-in-the-middle attacks, and a race condition in cloud-based device adoption — several of which can be chained together, and with two previously disclosed command-injection flaws, to compromise an organization’s entire fleet of managed network devices. Forescout found roughly 1,800 Omada controllers directly accessible from the internet despite vendor guidance against such exposure, and identified related weaknesses extending into TP-Link’s VIGI IP camera line, Festa routers, and its Tapo and Kasa smart home products. TP-Link has patched a portion of the disclosed issues, but says remediation for some of the more structural weaknesses won’t be complete until later in 2026, and certain low-severity findings won’t be patched at all.
Russian Hackers Abuse Hotel Wi-Fi Networks to Steal Microsoft 365 Credentials and Deploy Malware
Microsoft disclosed a campaign called CaptiveCrunch in which Storm-2945, a sub-cluster of the Russian state-linked group Midnight Blizzard, has spent months hijacking captive-portal Wi-Fi networks at hotels and conference centers worldwide to steal Microsoft 365 credentials and deploy malware against traveling corporate employees. Two malware strains anchor the operation — CornFlake, a full-featured Windows remote access trojan that masquerades as a legitimate Windows service and establishes multiple redundant persistence mechanisms, and ChocoShell, an information stealer built to quickly harvest browser cookies, saved passwords, and Microsoft 365 single sign-on tokens. Microsoft says the campaign shows signs of AI assistance and recommends business travelers avoid unexpected software update prompts on public networks, use a full-tunnel VPN so DNS traffic never touches untrusted hotel infrastructure, and treat any captive-portal network — including those at airports, universities, and conference venues — as fundamentally untrustworthy.
Hackers Steal Over $130M by Exploiting Bug in Offline Hardware Wallets
A software flaw in Coldcard hardware wallets, made by Canadian manufacturer Coinkite, has allowed at least 15 separate attackers to drain more than $130 million in Bitcoin from over 7,700 addresses since late July, despite the devices being marketed as “cold storage” that never connects to the internet. The bug traces back to a March 2021 firmware build that generated wallet seed phrases with far less randomness than advertised, making it mathematically feasible for attackers to reconstruct private keys offline — meaning victims who followed every standard security practice, including never sharing a seed phrase or connecting the device to a network, were still exposed. Coinkite has released patched firmware and is urging all affected users to migrate their funds to newly generated seed phrases immediately, while separately warning that AI-assisted code review failed to catch the underlying flaw — a caution the company says applies to any organization building security-critical hardware or software.
