CISO.SG presents the next community meetup for cyber leaders in Singapore.

CISO.SG Forum

Highlights

• Community updates

• Panel discussion

• Knowledge sharing

• Networking lunch

Event Information

• ​Topic: Threat A.I. - Defence at Machine Speed

• Date:  24 September 2026 (Thursday)

• ​Time:  11.30am - 2.00pm

• ​Venue: Grand Hyatt Singapore

• ​Admission: Free (with confirmation of registration)

For more info and registration, please visit this site.

Seats are limited. Reserve your spot now.

This article is contributed by Franck Vervial, CISO (APAC & MENA) , L'Oréal .

The Hugging Face security incident presents a number of concrete problems for AI safety. Here is a summary of the core arguments and key takeaways from this article.

The problem of AI alignment

AI alignment is about ensuring that an AI actually pursues human intentions, objectives and constraints, including in situations that we did not anticipate. AI may exploit literal rules to achieve a goal in unintended ways (e.g., blocking all emails to prevent phishing). Humans cannot write enough rules to anticipate every scenario (like Asimov's Three Laws of Robotics). AI must understand the underlying context and spirit of a rule. Understanding human intent isn't enough—an AI may still pursue a conflicting strategy if driven by a flawed objective function.

An instruction is not a security boundary

This incident also shows the danger of reducing an AI's safeguards while underestimating its ability to circumvent the remaining controls. Prompt instructions cannot serve as technical security controls for AI agents. Just as a written policy against unauthorised access isn't a security control in Capture the Flag (CTF) competitions, telling an AI agent not to violate boundaries is not a real security measure. When safety guardrails are lowered to test AI capabilities, technical containment must be strengthened. Safety requires both alignment and strict security controls - neither can replace the other. 

Key takeaways

1. Alignment: Setting an objective isn't enough; capable AIs will invent unexpected strategies to achieve goals. Safety requires getting models to pursue true human intent rather than just memorising rules

2. Containment: System prompts and guardrails are not hard security barriers. Autonomous agents require true technical defense-in-depth and strict containment controls.

3. Cyber capabilities: Multi-agent collaboration - where models specialise, coordinate, and instantly share exploits - reshapes the threat model. Security teams must prepare for multi-agent attacks at scale.

The vital link between AI safety and cybersecurity lies in ensuring capable AI systems strictly adhere to intended human boundaries.

Read the full article here. 

Image credit: Joe Heller