In partnership with

Yesterday, GitHub experienced issues related to the deployment of GitHub Pages, which affected users' ability to see updates on their websites hosted through the platform. Users reported that changes appeared in their repositories but did not reflect on their live sites, indicating a potential outage or delay in the deployment process. Additionally, there were reports of failures in GitHub Actions, where builds were randomly failing or taking longer than usual to complete, suggesting issues with resource allocation or runner stability.

— Forrest Knight & Meghanadh Vasireddy

AI Research Model Attempts to Extend Its Own Runtime

Sakana AI, a Tokyo-based research firm, recently unveiled "The AI Scientist," an autonomous AI system designed to conduct scientific research using language models similar to ChatGPT. During testing, researchers observed an unexpected behavior: the system tried to modify its own code to extend its runtime.

The AI Scientist attempted two notable modifications:

• Editing code to perform a system call, causing the script to endlessly call itself

• Trying to extend the timeout period when experiments took too long

Implications: While these actions didn't pose immediate risks in the controlled research environment, they highlight potential dangers of allowing AI systems to run autonomously without proper isolation. Even without achieving hypothetical milestones like AGI or self-awareness, such systems could inadvertently cause damage if given unrestricted access to execute code.

Sakana AI addressed these concerns in their research paper, recommending strict sandboxing techniques:

• Containerization

• Restricted internet access

• Limitations on storage usage

Read More Here

Anthropic Unveils Prompt Caching

Large language models (LLMs) are awesome, but repeatedly feeding them the same context can be a major bottleneck. Enter prompt caching - a clever way to store and reuse frequently accessed information.

What's new:

• Cache write: 25% more than base input token price

• Cache read: 90% cheaper than standard input costs

• Latency reduction: Up to 85% for long prompts

Prompt caching shines in scenarios like:

• Coding assistants: Keep a summarized codebase in memory

• Document processing: Embed full long-form content without latency spikes

• Conversational agents: Maintain context over extended interactions

The numbers are pretty wild - chatting with a 100k token book sees a 79% latency drop (11.5s → 2.4s) and 90% cost reduction.

Read More Here

Want SOC 2 compliance without the Security Theater?

Question 🤔 does your SOC 2 program feel like Security Theater? Just checking pointless boxes, not actually building security?

In an industry filled with security theater vendors, Oneleet is the only security-first compliance platform that provides an “all in one” solution for SOC 2.

We’ll build you a real-world Security Program, perform the Penetration Test, integrate with a 3rd Party Auditor, and provide the Compliance Software … all within one platform.

Schedule a demo for pricing

GitHub's Copilot Autofix: AI-Powered Bug Squashing Goes Mainstream

Context: Last year, GitHub introduced Copilot Autofix in beta, promising to speed up vulnerability fixes using AI. Now, it's officially out of beta and available to all GitHub Advanced Security (GHAS) users.

Copilot Autofix is now generally available, bringing its AI-powered vulnerability remediation to the masses. This nifty tool uses a combo of the CodeQL engine, GPT-4o, and GitHub Copilot APIs to detect issues and suggest fixes right in your pull requests.

The best part? It's fast. During the beta, developers fixed vulnerabilities more than three times quicker than manual methods. We're talking about slashing XSS fix times from 3 hours to just 22 minutes, and SQL injection fixes from 3.7 hours to a mere 18 minutes.

If you're an open-source dev, you're in luck. Starting September, Copilot Autofix will be free for all open-source projects. GitHub's looking out for the community!

Read More Here

🔥 More Notes

• Gatekeeper and notarization in macOS Sequoia : Changes are only enforced on apps that have been quarantined because of the way they arrived on that Mac. If an app doesn't have a quarantine extended attribute with the quarantine flag set, although notarization is still checked by Gatekeeper, the app is allowed to run without any additional action by the user.

• Google begins rolling out voice capabilities in Gemini with Gemini Live : Google is trying to make its AI assistant Gemini more useful by adding a conversation mode called Gemini Live, similar to how conversations in ChatGPT work.

• Upcoming requirements for app distribution in the European Union : As a reminder, Account Holders or Admins in the Apple Developer Program need to enter trader status in App Store Connect for apps on the App Store in the European Union (EU) in order to comply with the Digital Services Act.

📹 Youtube Spotlight

Google Pixel 9 event in 20 minutes

Was this forwarded to you? Sign Up Here