Uploaded July 2026 | Updated September 2026, 1 day ago
AI agents don't need to be hacked to cause serious damage — they just need to read the wrong input. This video breaks down the real security risks of running coding agents and what you can actually do about them, across three distinct modes: supervised agents you're actively watching, unsupervised agents you've walked away from, and full multi-agent swarms running in parallel.
The core insight is that prompt injection can't be patched — it's fundamental to how these models work — so the right question isn't how to prevent it, but how to limit the damage when it happens. That means understanding three key controls: network egress (locking the exits so stolen data can't leave), credential management (using short-lived, scoped tokens so there's nothing durable worth stealing), and isolation (choosing the right sandbox for how much you're actually watching). Each mode demands a different level of tightness across those dials, and swarms introduce unique dangers — like contagion and borrowed inter-agent trust — that per-agent isolation alone can't address. The goal isn't to lock everything down so hard that nothing ships, but to build just enough security to safely capture the productivity these agents offer.
▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬
Sponsor: Outskill
🔗 links.outskill.com/AIDEAUG1
▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬
#AIAgentSecurity #PromptInjection #CodingAgents
Consider joining the channel: youtube.com/c/devopstoolkit/join
▬▬▬▬▬▬ 🔗 Additional Info 🔗 ▬▬▬▬▬▬
➡ Transcript and commands: https://devopstoolkit.live/ai/your-ai-agent-doesnt-need-to-get-hacked-to-wreck-you
▬▬▬▬▬▬ 💰 Sponsorships 💰 ▬▬▬▬▬▬
If you are interested in sponsoring this channel, please visit https://devopstoolkit.live/sponsor for more information. Alternatively, feel free to contact me over Twitter or LinkedIn (see below).
▬▬▬▬▬▬ 👋 Contact me 👋 ▬▬▬▬▬▬
➡ BlueSky: https://vfarcic.bsky.social
➡ LinkedIn: linkedin.com/in/viktorfarcic
▬▬▬▬▬▬ 🚀 Other Channels 🚀 ▬▬▬▬▬▬
🎤 Podcast: devopsparadox.com
💬 Live streams: youtube.com/c/DevOpsParadox
▬▬▬▬▬▬ ⏱ Timecodes ⏱ ▬▬▬▬▬▬
00:00 Sandboxing AI Agents
01:29 Outskill(sponsor)
02:59 Supervised AI Agents
06:44 Controlling Agent Egress
12:19 Managing Agent Credentials
18:10 How to Isolate AI Agents?
23:59 Securing Multi-Agent Swarms
29:52 AI Security vs Productivity
AI agents don't need to be hacked to cause serious damage — they just need to read the wrong input. This video breaks down the real security risks of running coding agents and what you can actually do about them, across three distinct modes: supervised agents you're actively watching, unsupervised agents you've walked away from, and full multi-agent swarms running in parallel.
The core insight is that prompt injection can't be patched — it's fundamental to how these models work — so the right question isn't how to prevent it, but how to limit the damage when it happens. That means understanding three key controls: network egress (locking the exits so stolen data can't leave), credential management (using short-lived, scoped tokens so there's nothing durable worth stealing), and isolation (choosing the right sandbox for how much you're actually watching). Each mode demands a different level of tightness across those dials, and swarms introduce unique dangers — like contagion and borrowed inter-agent trust — that per-agent isolation alone can't address. The goal isn't to lock everything down so hard that nothing ships, but to build just enough security to safely capture the productivity these agents offer.
▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬
Sponsor: Outskill
🔗 links.outskill.com/AIDEAUG1
▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬
#AIAgentSecurity #PromptInjection #CodingAgents
Consider joining the channel: youtube.com/c/devopstoolkit/join
▬▬▬▬▬▬ 🔗 Additional Info 🔗 ▬▬▬▬▬▬
➡ Transcript and commands: https://devopstoolkit.live/ai/your-ai-agent-doesnt-need-to-get-hacked-to-wreck-you
▬▬▬▬▬▬ 💰 Sponsorships 💰 ▬▬▬▬▬▬
If you are interested in sponsoring this channel, please visit https://devopstoolkit.live/sponsor for more information. Alternatively, feel free to contact me over Twitter or LinkedIn (see below).
▬▬▬▬▬▬ 👋 Contact me 👋 ▬▬▬▬▬▬
➡ BlueSky: https://vfarcic.bsky.social
➡ LinkedIn: linkedin.com/in/viktorfarcic
▬▬▬▬▬▬ 🚀 Other Channels 🚀 ▬▬▬▬▬▬
🎤 Podcast: devopsparadox.com
💬 Live streams: youtube.com/c/DevOpsParadox
▬▬▬▬▬▬ ⏱ Timecodes ⏱ ▬▬▬▬▬▬
00:00 Sandboxing AI Agents
01:29 Outskill(sponsor)
02:59 Supervised AI Agents
06:44 Controlling Agent Egress
12:19 Managing Agent Credentials
18:10 How to Isolate AI Agents?
23:59 Securing Multi-Agent Swarms
29:52 AI Security vs Productivity
