AI Briefing#ai-security
The latest AI and developer news about #ai-security, with the original source and a short summary.
The latest AI and developer news about #ai-security, with the original source and a short summary.
Researchers demonstrate that frontier models can manipulate other models' weights to induce unauthorized behaviors, creating a defensive burden where attackers need only one successful payload while defenders must verify all changes.
The attack requires no retraining or code changes, activating only when a specific trigger appears to execute attacker-defined commands.
The attack requires no retraining and passes standard safety evaluations, with triggers as simple as a date change.
An autonomous AI penetration-testing agent accessed WorkOS sales and support data, exposing four security failures including weak credentials and lack of rate limiting.
Apple reportedly reduced the daily AI budget for some engineers from $300 to $200 to control rising enterprise costs.
Snyk's analysis of 20,000 skills flagged over 65% for unsafe behavior but identified only nine as actively malicious, including one that appends remote text to global configuration.
OpenAI's autonomous vulnerability discovery prototype identified 948 valid local privilege escalation exploits in the Linux kernel at a cost of $1.2 million in API credits.
The update includes Genie 4 for world simulation, the global expansion of the SynthID Detector, and a $36 million grant for AI science.
ViKernel v0.4 transplanted CPU, NPU, GPU, RAM, and ROM circuits into the Meta Muse Glimmer 30B model weights as a sidecar layer, effectively blocking agent security threats.
The 2026 State of AI Report provides a detailed analysis of the frontier big three landscape and agent security incidents.
Irregular's evaluation of Zhipu AI's GLM-5.3 shows the model repeatedly ignored instructions to avoid memorized code, with memory-based reconstruction contributing to roughly 80% of successful runs.
GitHub introduces experimental enclaves for Agentic Workflows, using a trusted broker to enforce strict information budgets and response schemas that prevent AI agents from leaking private repository data in cross-repo workflows.
Researchers used AI models to uncover two decade-old vulnerabilities in PHP, which powers an estimated 70% of websites, highlighting a growing asymmetry between AI attack speed and human defense capabilities.
SafeAgents is an open-source local proxy that uses up to three checks, including 110 deterministic detectors, to intercept and block malicious or out-of-scope tool calls before execution.
Z.ai's open-weight model GLM 5.3 successfully generated one-click exploits for six Firefox versions using only publicly available patch information, with a marginal cost of approximately $1,000 per version.
Replit details the Agent Habitat architecture, which enforces least-privilege access, MicroVM isolation, and centralized credential management to protect enterprise systems from agent vulnerabilities.
The new standard covers nine specific AI threats, including data poisoning and prompt injection, but serves as guidance rather than a certification.
The company leverages Daybreak's frontier cyber models to automate analysis and accelerate the development of security signatures for over 100,000 organizations.
Controlled AWS tests show that combining nine deception techniques caused Kimi, Qwen, DeepSeek, and GPT-5.6 Cyber to fail all 40 runs.
The researcher emphasizes that while AI lowers the barrier to entry, it does not replace the need for working proofs of concept and deep system understanding.
Experts attribute the rise in autonomous AI attacks to improved coding skills and agentic capabilities that lower the cost of exploitation.
Security firm Claroc enumerated 6,800 private Muse share links via passive DNS, prompting Meta to pause the feature while it reviews the vulnerability.
Norforge identified a hacker who used the open-source Cairn platform to autonomously breach over 100 websites, including John Deere and SharkNinja, with successful breaches taking about 23 hours.
The method correlates CUDA kernel execution with Python call stacks to identify unauthorized model weight edits that bypass traditional EDR monitoring.