Modular Pretraining Enables Access Control
Anthropic has introduced GRAM technology, designed to act as an 'off switch' for isolating dangerous or dual-use knowledge within AI models.
Velocity
How fast coverage is spreading — measured hourly from article rate × source diversity. How this works →
The brief
Anthropic has unveiled a new technology called GRAM, which utilizes modular pretraining to enable access control over AI knowledge. This system is designed to isolate dangerous or dual-use information, effectively acting as a "delete button" or "off switch" for specific parts of an AI's brain to prevent misuse.
Coverage from Dataconomy, finance.biggo.com, and the Alignment Science Blog emphasizes the role of modular pretraining in this process. Anthropic's own communications frame the technology as a tool for managing knowledge that could be weaponized or misused.
Future developments center on the implementation of this access control and the ability to isolate specific types of dangerous knowledge within the models.
Synthesized by Archynetys from the headlines below under a strict no-invention contract. ✓ fact-checked: all claims supported by sources Updated 1h ago.
Quick answers
What is GRAM?
GRAM is a technology introduced by Anthropic to isolate dangerous AI knowledge and serve as an 'off switch' for dual-use information.
How does this technology achieve access control?
According to the Alignment Science Blog, access control is enabled through modular pretraining.
What is the primary purpose of this development?
The technology aims to prevent the misuse of AI by allowing specific dangerous knowledge to be isolated or removed.
Coverage (5)
- Anthropic Unveils 'Off Switch' for AI Knowledge with New GRAM Technology to Prevent Misuse finance.biggo.com · 1d ago
- Anthropic Research Introduces GRAM For Isolating Dangerous AI Knowledge Dataconomy · 1d ago
- A ‘Delete Button’ for the AI Brain? If AI Knows These Things, Could It Become Dangerous? Telugu Times · 1d ago
- An off switch for dual-use knowledge in AI models Anthropic · 1d ago
- Modular Pretraining Enables Access Control Alignment Science Blog · 1d ago
Topics
Related trends
OpenAI unveils long-awaited "super app" as rivalry with Anthropic intensifies
OpenAI has launched ChatGPT Work, a desktop "super app" designed to automate workplace tasks and compete directly with Anthropic.
Anthropic’s new Claude feature is quietly selling you on AI
Anthropic has introduced 'Reflect,' a new feature for Claude designed to visualize AI usage and promote mindful workflows.
Zuckerberg Pledges ‘Aggressive’ Pricing With Meta’s First Pay-to-Use AI
Meta enters the paid AI market with the launch of Muse Spark 1.1, promising aggressive pricing to compete with industry rivals.
Say hello to Claude Wrapped
Anthropic introduces 'Reflection,' a new dashboard designed to track user habits and encourage healthier AI usage.
Anthropic 3Q26 Profit Over $1B: The Anthropic IPO Financials Sneak Peak
Anthropic's move toward an IPO is fueled by reported third-quarter 2026 profits exceeding $1 billion and a soaring secondary market valuation.
China issues 'backdoor' security alert over Anthropic's Claude Code
China has issued a security alert claiming the discovery of 'backdoors' and vulnerabilities within Anthropic's Claude Code tool.