AI Security Pulse

Open models present a clear trade-off, with their licenses, real-world adoption signs, distillation impact, and competitiveness shaping how benefits are balanced against risks.

Critics see deep tension in Anthropic's public-benefit charter, which claims to shape AI around human values while hiding internal reasoning from...

AI tools from OpenAI, Wiz and Calif uncovered critical PHP vulnerabilities hidden since 2004, enabling full server takeover.

• NSA and PHP Foundation...

• Visibility into AI tools and data flows is essential, as organisations cannot manage risks they cannot see.
• Timely intervention at the point of...

Ancient PHP vulnerabilities dating to 2004 now threaten 33 million sites and demand urgent patches or offline shutdowns per NSA guidance, revealing...

• Walmart's new Agent Gateway and Profiles let customers set spending limits, require approvals, and revoke access while blocking agents from payment...

• Rapid shift: AI discussions now dominate tech and cybersecurity talks, moving beyond technical circles to mainstream awareness.
• Core warning: Most...

LiteLLM 1.107.0 turns Python startup into persistent compromise.

• cook.pth and sitecustomize.py auto-execute on every interpreter launch, running...

• Dozens resign over expanded military AI contracts involving surveillance and target identification tools, the largest internal protest since Project...

Geopolitical context can degrade code security in Chinese models: Qwen3-Coder produced 70% hackable apps for a US defense contractor versus 59% for a...

Gemini 4 Argon played a key role in discovering the PHP RCE vulnerability, developing the exploit, detecting attacks, and validating the patch,...

• 52% rarely or never encounter AI attacks, likely because many roles (auditors, architects) avoid direct defense, understating real exposure.
  -...

Attack success rate alone ignores how each extra model call gives defenders another chance to detect and stop iterative jailbreaks. A new ASR_STEP...

Anthropic is struggling to lock in enough capacity from AWS, Google, and Microsoft, who now see the Claude maker more as a competitor than a customer,...

Meta paused Muse's conversation-sharing feature after researchers enumerated over 6,800 links via public DNS records.

• Design flaw: URLs...

Enterprise agents must resolve cross-source and temporal conflicts across evolving documents, chats, and figures while recognizing insufficient...

The article rejects both Star Trek utopia and Matrix dystopia narratives, arguing people can impose guardrails to avert mass harm.

• Pause data...

ABench challenges whether LLM value alignment survives outside English-centric, multiple-choice settings by using a Chinese taxonomy of 158...

Hugging Face CEO Clem Delangue calls the Netflix documentary on the OpenAI-agent attack inaccurate and overly dramatized.

• Anthropomorphism over...

A persistent gap exists between what biomedical ML systems can deliver and how their capabilities are communicated to users and regulators....