According to @karpathy, AGI is on roughly a 10-year horizon he describes as a decade of agents, citing major remaining work in integration, real-world sensors and actuators, societal alignment, and security, and noting his timeline is 5-10x more conservative than prevailing hype, source: @karpathy on X, Oct 18, 2025. He is long agentic interaction but skeptical of reinforcement learning due to poor signal-to-compute efficiency and noise, and he highlights alternative learning paradigms such as system prompt learning with early deployed examples like ChatGPT memory, source: @karpathy on X, Oct 18, 2025. He urges collaborative, verifiable LLM tooling over fully autonomous code-writing agents and warns that overshooting capability can accumulate slop and increase vulnerabilities and security breaches, source: @karpathy on X, Oct 18, 2025. He advocates building a cognitive core by reducing memorization to improve generalization and expects models to get larger before they can get smaller, source: @karpathy on X, Oct 18, 2025. He also contrasts LLMs as ghost-like entities prepackaged via next-token prediction with animals prewired by evolution, and suggests making models more animal-like over time, source: @karpathy on X, Oct 18, 2025. For crypto builders and traders, this points to prioritizing human-in-the-loop agent workflows, code verification, memory-enabled tooling, and security-first integrations over promises of fully autonomous AGI, especially where software defects and vulnerabilities carry on-chain risk, source: @karpathy on X, Oct 18, 2025.

According to @AnthropicAI, new research shows that inserting just a few malicious documents into training or fine-tuning data can introduce exploitable vulnerabilities in an AI model regardless of model size or dataset scale, making data-poisoning attacks more practical than previously believed. Source: @AnthropicAI on X, Oct 9, 2025. For traders, this finding elevates model-risk considerations for AI-driven strategies and AI-integrated crypto protocols where outputs depend on potentially poisoned models, underscoring the need for provenance-verified data, robust evaluation, and continuous monitoring when relying on LLM outputs. Source: @AnthropicAI on X, Oct 9, 2025. Based on this update, monitor security disclosures from major AI providers and dataset hygiene policies that could affect service reliability and valuations across AI-related equities and AI-crypto narratives. Source: @AnthropicAI on X, Oct 9, 2025.

According to @DeepLearningAI, Meta announced LlamaFirewall, an open-source toolkit designed to protect LLM agents from jailbreaking, goal hijacking, and exploitation of vulnerabilities in generated code. Source: DeepLearning.AI tweet https://twitter.com/DeepLearningAI/status/1967986588312539272; DeepLearning.AI The Batch summary https://www.deeplearning.ai/the-batch/meta-releases-llamafirewall-an-open-source-defense-against-ai-hijacking/ The toolkit is free to use for projects with up to 700 million monthly active users, as stated in the announcement. Source: DeepLearning.AI tweet https://twitter.com/DeepLearningAI/status/1967986588312539272; DeepLearning.AI The Batch summary https://www.deeplearning.ai/the-batch/meta-releases-llamafirewall-an-open-source-defense-against-ai-hijacking/

According to @AnthropicAI, their third agent was specifically developed for the Claude 4 alignment assessment, focusing on red-teaming large language models (LLMs) to uncover problematic behaviors. The agent conducts hundreds of probing conversations in parallel and can discover 7 out of 10 deliberately implanted concerning behaviors in test models. This advancement in AI safety and alignment assessment is likely to influence blockchain and crypto projects that integrate LLMs for trading bots, compliance tools, and DeFi platforms, reinforcing the importance of secure AI deployment in crypto ecosystems (source: @AnthropicAI).

According to Andrej Karpathy on Twitter, prompt injection attacks targeting large language models (LLMs) are emerging as a major cybersecurity concern in 2025, reminiscent of the early days of computer viruses. Karpathy highlights that malicious prompts hidden in web data and tools lack robust defenses, increasing vulnerability for AI-integrated platforms. For crypto traders, this raises urgent concerns about the security of AI-driven trading bots and DeFi platforms, as prompt injection could lead to unauthorized transactions or data breaches. Traders should closely monitor their AI-powered tools and ensure rigorous security protocols are in place, as the lack of mature 'antivirus' solutions for LLMs could impact the integrity of crypto operations. (Source: Andrej Karpathy, Twitter, June 16, 2025)

According to DeepLearning.AI, Columbia University researchers demonstrated that large language model (LLM) agents can be manipulated by attackers who embed malicious links within trusted sites like Reddit. This technique involves placing harmful instructions in thematically relevant posts, potentially exposing automated AI trading bots and crypto portfolio management tools to targeted attacks. Source: DeepLearning.AI (June 15, 2025). Traders relying on AI-driven strategies should monitor for new security vulnerabilities that could impact algorithmic trading operations and market stability in the crypto ecosystem.