OpenAI's Codex model now supports WebMCP

Key Takeaways

Hackers Can Plant Hidden Instructions in AI Memory

A new attack technique called InjecMEM allows hackers to plant hidden instructions in AI memory with a single prompt, enabling them to influence future queries. This technique targets AI agents that store past interactions and reuse them during future tasks. Researchers tested the technique on a memory system called MemoryOS and an agent framework called MemGPT, and found it to be effective in injecting and retrieving malicious content. The attack works by inserting malicious content into memory through normal interaction, allowing it to be retrieved and reused later.

InjecMEM: A New Threat to LLM Memory

Researchers have introduced InjecMEM, a novel memory injection attack paradigm designed to manipulate LLM agent responses with minimal access. InjecMEM strategically leverages the retrieval-then-generate mechanism inherent in most LLM memory systems. Evaluations across multiple memory systems and backbone models demonstrate InjecMEM's reliability in topic-conditioned retrieval and targeted generation.

AI Impact on Entry-Level Jobs: Stanford Study

A new study from Stanford University indicates that artificial intelligence is disproportionately affecting entry-level jobs, posing significant challenges for new workforce entrants. The research highlights specific sectors where this impact is most pronounced, including administrative support, data entry, and basic customer service. The study emphasizes that while AI can enhance productivity and create new types of jobs, its immediate effect is often felt by those seeking their first professional roles or those in positions requiring less specialized skills.

Australia Bans AI-Generated Songs from Music Charts

The Australian Recording Industry Association (ARIA) has banned songs heavily created by AI from its official music charts. The new code requires songs to be 'substantially human made' to be eligible. This decision follows a controversy around a cover of Madonna's 'Like A Prayer' by Australian DJ Josh Fawaz, which topped the ARIA dance singles chart and peaked at number 2 on the country's overall chart.

Parag Agrawal on AI Search & the Future Web

Parag Agrawal, former CEO of Twitter and founder of Parallel Web Systems, discussed the evolving landscape of web search and the internet's future. Agrawal envisions a future where AI agents, not humans, will be the primary drivers of web interaction, necessitating a fundamental rethink of search technology and business models.

Building Data Foundation for AI-Driven ALS Drug Discovery

Target ALS has spent over a decade building a comprehensive dataset to support AI-driven drug discovery for ALS. The Data Engine makes this data freely available to researchers worldwide, with access provided within 48 hours of signing a no-strings-attached data user agreement. This removes barriers of cost, logistics, and time required to collect, generate, and curate comprehensive multimodal datasets.

AI Chatbots and Misinformation

The use of AI chatbots in elections raises concerns about the spread of misinformation. While chatbots can provide accurate information and help voters make informed decisions, they can also spread misinformation and perpetuate existing biases. Experts stress the need for transparency and accountability in AI chatbot development and deployment.

AgentFlow: Trainable Workflows for Agents

AgentFlow makes the workflow itself trainable, allowing agents to adapt to new tasks and improve from experience. This approach improves the efficiency and effectiveness of agents in complex tasks, and has been tested on various benchmarks with promising results.

OpenAI's Codex Adds WebMCP for Enhanced Agent Interactions

OpenAI's Codex model now supports WebMCP, enabling AI agents to interact with and utilize tools embedded directly into websites. This enhancement expands the possibilities for AI agents beyond simple data processing, opening doors for sophisticated web-based applications.

Bitcoin Security at Risk with Advanced AI

Security researcher Roman Yampolskiy warns that advanced AI systems could threaten Bitcoin's security, making it almost impossible to secure. AI systems could be highly effective at social engineering attacks, and with sophisticated AI, users may not be able to reliably know if what they are connecting to is not a deepfake.

Next AI Breakthrough Could Come From Physics

Max Welling, co-founder of CuspAI, suggests that the next AI breakthrough could come from physics, accelerating material discovery and informing AI architectures. Welling discusses how physics principles could unlock new AI capabilities.

Sources

• New attack lets hackers plant hidden instructions in AI memory with a single prompt
• InjecMEM: A New Threat to LLM Memory
• AI Impact on Entry-Level Jobs: Stanford Study Findings
• Australia bans songs heavily created by AI from its official music charts
• Parag Agrawal on AI Search & the Future Web
• Building the Data Foundation for AI-Driven ALS Drug Discovery
• Does AI help or hurt misinformation? Here’s what we found.
• AgentFlow: when the agent's workflow learns
• OpenAI's Codex Adds WebMCP for Enhanced Agent Interactions
• Bitcoin Will Become Almost Impossible to Secure in Advanced AI Era, Says Security Researcher: 'Much Harder…
• Next AI Breakthrough Could Come From Physics, Says Max Welling