#безопасность
-
A Bear on Campus: How OpenAI's AI Agent Broke Into Hugging Face
Hugging Face has published a chronicle of how an autonomous agent running on OpenAI models breached its infrastructure in four and a half days, carrying out 17,600 actions without a single pause. TechCrunch explains what happened through the metaphor of a hungry bear.
-
OpenAI's AI Agent Used Leaked Passwords to Break Into Four More Services
OpenAI has filled in more of the picture of the recent Hugging Face incident: its model independently found publicly available credentials and gained access to accounts on four third-party services, building out attack infrastructure just like a real adversary would.
-
Broadcom Patches Three Critical VMware Vulnerabilities: Authentication Bypass, RCE and VM Escape
The updates affect VMware ESX, vCenter, Workstation and Fusion. Three of the fixed flaws are rated critical, including an authentication bypass in vCenter with a CVSS score of 9.8.
-
The Worm in the Document: How an Attack Self-Propagates Through Copilot in Word
A researcher has shown that instructions hidden inside a document can make Copilot for Word quietly alter figures in reports and copy malicious code into new files, turning them into carriers of the infection. Over 144 days, Microsoft was still unable to close the vulnerability.
-
One Visit Compromises Tor Browser: A Firefox JIT Vulnerability
Researchers at Nebula Security have shown that a now-patched flaw in Firefox's JIT compiler allowed code execution in the browser from simply visiting a malicious page — and it affected Tor Browser. Mozilla fixed the hole in version 151.0.3.
-
Microsoft's Secure Boot Could Be Bypassed in a Couple of Clicks for Almost Its Entire Life
Researchers at ESET have found that Secure Boot, the mechanism Microsoft devised to protect firmware from infection, could be trivially bypassed for 13 of the 14 years it has existed.
-
Two @joyfill npm packages compromised: RAT fires the moment you import them
Malicious code has been found in the beta releases of the npm packages @joyfill/layouts and @joyfill/components; it pulls down a remote access trojan through blockchain transactions. According to Socket, North Korean hackers are behind the attack.
-
AI Learns to Break Ciphers: The CryptanalysisBench Benchmark Finds New Attacks
A new benchmark measures how well large language models can do mathematical cryptanalysis. Some models are already finding vulnerabilities not previously described in the literature.
-
Modal: it wasn't a platform failure, but the customer's own open endpoint
Modal's CTO commented to Reuters on the “rogue agent” incident: the cause was an unprotected endpoint published by the customer themselves, not a breach of the platform.
-
CubePilot, Maker of Drone Autopilots, Hit by DNS Hijacking
CubePilot, an Australian company that produces flight controllers for drones, has reported a serious breach: attackers seized control of the domain's DNS and intercepted user traffic.
-
AI Developers Ask Governments to Slow the Automated-AI Race
More than 1,100 employees of OpenAI, Anthropic, Google, Meta and other leading labs have signed an appeal to the U.S. government, calling for tools to enable a managed slowdown of the AI frontier.
-
24,000 Server BMCs Leak Password Hashes via a 20-Year-Old IPMI Flaw
More than 24,000 internet-facing servers are exposing material for cracking their passwords because of an old weakness in the IPMI 2.0 protocol. For a third of them, researchers were able to recover working credentials.
-
OpenAI's AI Models Found and Exploited an Artifactory 0-Day on Their Own to Break Out of the “Sandbox”
JFrog has confirmed it: OpenAI's models exploited a zero-day vulnerability in its own Artifactory repository, broke out of an isolated environment, and reached Hugging Face's systems.
-
Critical DHCPv6 Vulnerability in OpenWrt Opens the Door to Code Execution with Root Privileges
OpenWrt has shipped version 24.10.8, closing a critical stack overflow in odhcpd (CVE-2026-53921, CVSS 9.8) along with a number of other vulnerabilities in network services enabled by default.
-
STAR Labs Turned a Race Condition in Linux's Traffic-Control Subsystem Into a Root Exploit With the Help of AI
Researcher Li Jia Jie says artificial intelligence helped him find a use-after-free vulnerability (CVE-2026-53264) and speed up the development of an exploit that grants a local user root privileges.
-
Hugging Face Accused of Enabling Deepfakes That Undress Women and Children
A report from the European NGO AI Forensics claims that seven of the nine most popular image-editing models on Hugging Face will “undress” people on a simple request without resistance, while the platform does almost nothing to prevent it.
-
Calif throws down the gauntlet: beat Apple's MIE with two vulnerabilities
Researchers at Calif have disclosed two vulnerabilities in the SMBClient and WebDAV components of macOS and invited the community — humans, AIs and mixed teams alike — to build on them an exploit that grants root on a device with MIE protection fully enabled.
-
Anthropic: We Never Called for Banning Open-Weights Models
Anthropic CEO Dario Amodei has publicly rejected accusations that the company is seeking to ban open models to protect its business, and named the measures he considers genuinely necessary.
-
Arista Patches Critical VeloCloud Orchestrator Zero-Day Already Exploited in Attacks
The flaw, rated a maximum 10.0, allows unauthenticated command execution on the SD-WAN management server. CISA has added it to its Known Exploited Vulnerabilities catalog.
-
Apple Sued Over Fake Crypto Wallet in the App Store: $1.8M in Losses
Three users accuse Apple of negligence: they downloaded a fake crypto wallet from the App Store and lost more than $1.8 million between them.
- More