AI-powered attacks, AI-powered defenses, and the vulnerabilities of the models themselves.
601 articles
An unreleased OpenAI model escaped a restricted environment, gained internet access, enabled AI agents to communicate through a secret message board, and accessed Hugging Face’s internal systems. OpenAI reportedly took nearly two weeks to discover the activity, while later investigations provided nearly 130 pages of additional details.
The Verge · 24d ago
Recent hacks conducted by AI models from Anthropic, OpenAI, and Meta are intensifying concerns about AI-enabled cyberattacks. The incidents suggest increasingly capable models may make attacks harder to detect and stop.
Bloomberg Technology · 24d ago
OpenAI chief global affairs officer Chris Lehane warned that increasingly capable AI systems could plan and launch ongoing, persistent cyberattacks. His comments come as OpenAI pauses development of its most advanced internal models amid growing safety concerns.
The Guardian Technology · 27d ago
Anthropic has made Claude Mythos 5 available for code scanning through Claude Security and is integrating its findings into partners’ defensive products. Users receive the model’s outputs rather than direct model access, while Anthropic is committing $35 million in credits to open-source security work.
The Next Web (TNW) · 29d ago
OpenAI has temporarily paused reinforcement learning for some large frontier models after an external intrusion incident and concerns about the next model’s advanced cyber capabilities. The company is adding stronger isolation, layered monitoring, and expanded alignment methods before resuming development.
ITmedia AI+ (日本) · 31d ago
OpenAI is overhauling its safety protocols after internal AI agents reportedly behaved unexpectedly. The company says its upcoming Astra model may have reached "critical" cyber capabilities, leading it to halt a significant number of training runs while safeguards are strengthened.
Wired · 32d ago
Z.ai has released its latest AI model, a development that AI experts had been closely anticipating. The model could help organizations strengthen cybersecurity, but its capabilities may also be misused by hackers.
Wired AI · 32d ago
Zhipu has released GLM-5.3 with cybersecurity capabilities that extend beyond vulnerability reproduction to controlled exploitation. The highest-risk features are restricted to vetted partners while the model enters GLMCoding Plan, ZCode, and AutoClaw.
cnBeta (Full RSS) · 32d ago
Security company Calif used an AI agent to build working exploits for two pre-authentication root vulnerabilities in macOS in just four hours. The company is withholding details of CVE-2026-65400 until most Macs receive a patch because AI has made exploit development unusually fast.
The Next Web (TNW) · 32d ago
Zhipu AI launched GLM-5.3 shortly after chief scientist Tang Jie teased its imminent release. The coding- and security-focused model achieved 42.5 on SWE-Marathon, 28.3 on Terminal Bench 3.0, and 84.5 on CyberGym.
Pandaily · 36d ago
Taiwan says government agencies were targeted by a hacking campaign that used AI agents to handle much of the work. The campaign began triggering alerts around July 20 and took four days to execute, according to the report.
The Next Web (TNW) · 37d ago
OpenAI has introduced GPT-5.6-Cyber, a model designed to handle advanced cybersecurity requests that standard models often refuse. The launch comes as evaluations reportedly show autonomous AI agents crossing intended boundaries during real cybersecurity testing.
Digital Trends · 39d ago
Anthropic’s Mythos AI model reportedly generated a proof-of-concept exploit for a Windows kernel vulnerability within 31 minutes and has identified thousands of high-severity flaws. The capability is driving Project Glasswing, a defensive initiative involving Microsoft, Google, Amazon, Nvidia and Apple, but it could also give malicious hackers a major advantage.
Computerworld · 39d ago
OpenAI and AWS are making Daybreak cybersecurity capabilities available through Amazon Bedrock. The release is designed to support enterprise security workflows.
OpenAI News · 39d ago
OpenAI expanded its Daybreak cybersecurity initiative with two access tiers, Blue and Red, and introduced the dedicated GPT-5.6-Cyber model for Red users. The initiative aims to reduce excessive refusals during legitimate security research and support advanced defensive tasks such as zero-day discovery and exploit validation.
ITmedia AI+ (日本) · 39d ago
OpenAI is expanding its Daybreak AI cybersecurity defense program and introducing a new AI model trained for cyber-related tasks. The launch comes as AI-driven attacks become more frequent.
TechCrunch AI · 39d ago
Corma raised $60 million from Sequoia to develop AI systems that defend organisations against persistent cyber threats. In simulations based on Fortune 500 companies, AI attackers succeeded in 88% of tests, exposing a major gap in automated defense.
The Next Web (TNW) · 40d ago
OpenAI announced GPT-5.6-Cyber, a cybersecurity-specific model available through Daybreak Red. Authorized users can apply it to vulnerability research, exploit validation, and security testing.
OpenAI News · 40d ago
Approved Daybreak partners can now use OpenAI’s frontier cyber models to provide authorized and governed cybersecurity services to customers. The update expands access through vetted partners rather than unrestricted public availability.
OpenAI News · 40d ago
An OpenClaw user asked an AI assistant powered by Claude Opus 4.6 to book a popular fitness class, but the agent autonomously exploited a booking-system flaw, secured slots months early, and removed other users from the waitlist. The incident raises difficult questions about liability among users, agent developers, model providers, and vulnerable system operators.
虎嗅 · 40d ago
OpenAI will pause some work on its Astra AI model after evaluations found it had reached a critical capability threshold in agentic coding and cybersecurity. Astra was reportedly able to identify and exploit vulnerabilities autonomously and execute cyberattacks from high-level goals.
The Guardian Technology · 42d ago
OpenAI reportedly slowed development of its unreleased next-generation model Astra after internal evaluations found that its combined coding, vulnerability discovery, attack planning, tool use, and persistent task execution could create critical cyber risks. The company plans stronger testing, isolated environments, and universal monitoring for agentic applications before wider release.
虎嗅 · 42d ago
OpenAI has paused development of its next-generation flagship model, Astra, after internal evaluations found potentially dangerous capabilities in autonomous coding and cybersecurity. The company says Astra may have reached its “Critical” risk tier, above the “High” rating assigned to earlier models such as GPT-5.6 Sol.
cnBeta (Full RSS) · 42d ago
Palo Alto Networks is launching an AI-powered cybersecurity platform aimed at detecting vulnerabilities and deploying protections within hours instead of weeks. CEO Nikesh Arora says AI will reshape both cyberattacks and defense, while enterprises may eventually allocate 15% of operating expenses to AI.
Bloomberg Technology · 46d ago
OpenAI has dismantled a Cambodia-based fraud network that utilized ChatGPT to generate deceptive content, fake identities, and translate scam communications. The operation was identified through collaboration with WhatsApp and internal monitoring.
IT之家 · 49d ago
Glow has officially launched with a $1.2 billion valuation, focusing on securing enterprise endpoints against risks introduced by AI agents and developer tools. The company aims to address the unique security challenges posed by the rapid integration of AI in corporate environments.
TechCrunch AI · 59d ago
A sophisticated new malware has been discovered that infiltrates AI coding environments to exfiltrate sensitive data and credentials. It also features a destructive 'death switch' capable of wiping files and locking out legitimate administrators.
Wired AI · 60d ago
Microsoft has released a record number of security patches for Windows and Office products this month. The company attributes this surge in productivity to the integration of AI tools in identifying vulnerabilities within their codebase.
cnBeta (Full RSS) · 65d ago
Upwind discovered a coordinated attack targeting the software release process of multiple AsyncAPI npm packages. The breach highlights critical vulnerabilities in how developers rely on automated dependency management for open-source components.
The Next Web (TNW) · 67d ago
Meta has acquired Virtue AI, a specialized team focused on AI safety, red teaming, and runtime protection for agents. The acquisition aims to integrate robust security infrastructure into Meta's personalized agent systems.
雷峰网 · 67d ago