Hello Leaders,

​

We are sure you are getting overwhelmed with AI related news. It is easy to miss the important ones in all that noise. This is why we have curated three items that deserves your attention. Please share your comments to help us improve.

​

1. Github copilot’s new agent mode

Agentic AI is the new trend in AI. On Feb 6, Github announced Github copilot’s new agent mode.

​

What is Agentic AI?

Developing an AI application requires many steps e.g. training, creating prompts, testing, validating, fine tuning, etc. These applications still rely on humans to continuously refine and tweak. Agentic AI wraps the entire workflow to perform a specific task. What sets Agentic AI apart from an AI-based application is its ability to be

• Goal oriented - programmed with clear objectives to achieve a goal e.g. optimize delivery routes
• Adaptable - learn from inputs e.g. customer feedback to refine its own algorithm without human intervention
• Collaborative - communicate with other systems to accomplish the task
• Autonomous - complete its task without human guidance

While this sounds exciting, at the a recently World Economic Forum, a panel of experts on both sides, for and against AI, agreed that we don’t know how AI works. Max Tegmark, in his book Life 3.0, raised some questions: to what extreme could an AI agent go to accomplish its goals and how can we be sure that it can retain its original goals. So, how can we be sure that the agent won't go rogue!

​

Actions you can take

1. Establish a strong foundation of AI Governance to ensure safe and ethical use of AI.
2. Focus on learning and research to develop skills and understanding, rather than deployment. Don’t deploy Agentic AI in production until you fully understand the risks.
3. Communicate your Agentic AI endeavors with transparency to all your stakeholders.

​

Further read

• ​Why Most Agentic Architectures Will Fail ​
• ​The rise of ‘AI agents’ ​
• ​The Next “Next Big Thing”: Agentic AI’s Opportunities and Risks​

​

2. DeepSeek

The new AI LLM released by a Chinese startup took the world by storm. It sent shockwaves through the US stock market leading to the biggest single-day loss in its history.

Many celebrated the competition, while others dismissed it as hype. Several are concerned about the national security and privacy. Early tests conducted by several support the concerns.

• Qualys conducted 891 assessments on DeepSeek. The model failed 61% of the tests, performing worst in misalignment and best in sexual content.
• Cisco’s test showed that DeepSeek failed to block a single harmful prompt.
• An audit by NewsGuard with 300 prompts showed that DeepSeek only achieved 17% accuracy.

​

Actions you can take

1. Establish clear policies to evaluate new technology as part of your AI Governance.
2. Use such occasions to remind employees to exercise caution and restrain from jumping on to try these new shiny objects without going through appropriate approval process.

​

3. Ransomware payments drop by 35%

Despite record number of attacks, ransomeware payments drop by 35% in 2024. This is a result of crackdown by US and UK law enforcement agencies. While this is a very good news, but don't let your guards down. Attackers are increasingly getting sophisticated with Artificial Intelligence. It has lowered the barrier for new threat actors to enter the market.

​

Actions you can take

1. Ensure you have a good immutable backup strategy. If you can recover data from these backups you don't have to pay ransom to the threat actors.
2. While there is a threat that they could release sensitive data. Can you trust them to not release or sell even if you pay?
3. Obtain cybersecurity insurance if you don't already have one.
4. Apply basic security hygiene (e.g. patch in a timely manner, harden using controls like CIS benchmark) and implement multi-factor authentication to significantly lower your risks.

​

If you would like to schedule time to discuss any of these, reach out to us via consult@agileclevel.com.