Video summary
Bezpieczeństwo w pracy z AI: automatyzacje i agenty
Main summary
Key takeaways
Main ideas & lessons (by section)
Overall goal of the lesson
- The speaker answers common safety questions from participants building work automation with AI agents.
- Core message: use proven “good practices” to reduce risks; it’s okay to revisit later once you begin building.
Methodologies / instruction lists
1) Where to run AI-agent orchestration tools (e.g., n8n, Make) + securing API keys
- Prefer self-hosted orchestration on your own infrastructure (if possible)
- Example: n8n self-hosted to reduce exposure to attackers.
- If self-hosting, you must manage server security
- Regular updates
- Proper configuration
- Harden the server against bots/scanners (e.g., hide from internet vulnerability scanning)
- If you lack admin/security experience, it may be better to use ready-made solutions.
- Analyze how the tool handles “secrets”
- Do not store API keys/passwords as plain text.
- Use the tool’s dedicated secrets modules (n8n and Make have these).
- Prevent secrets from leaking via logs/UI/errors
- Secrets should not appear in:
- logs
- graphical interfaces
- error messages sent to channels (e.g., Slack developer channels)
- Consider adding a log/error cleanup step (node) to scrub secrets.
- Secrets should not appear in:
- Use external secret managers (more professional approach)
- Examples: HashiCorp Vault, Google Secret Manager, AES Secret Manager, Volt (and similar).
- Automation should access secrets only when needed.
- Apply least privilege
- When cloud services connect to your internal DB:
- Use provider documentation to get IP address ranges
- Restrict firewall access to:
- those IP ranges
- other trusted addresses (e.g., your home IP)
- Goal: don’t expose internal “toys” to unrestricted internet access.
- Secure connectivity alternatives:
- Cloudflare tunnels
- Tailscale
- WireGuard
- Additional least-privilege guidance:
- Don’t let models perform destructive actions as superuser/admin.
- Create a dedicated user with limited permissions.
- Avoid running services with overly high privileges without proper hardening.
- When cloud services connect to your internal DB:
- Use per-project API keys
- Prefer per-project keys over a single global key.
- If supported, enforce extra restrictions:
- cost limits
- limits to specific endpoints/models (and “handpoints”)
2) Preventing “huge bills” from a paid AI API (token/cost draining)
- This class of attack is called denial-of-wallet (wallet drain).
- Note: attackers may need few or even one clever query that triggers an agent loop.
- Use defense in depth (multi-layer controls):
- Provider-level limits
- Set daily limits per API key or per model.
- Enable SMS/email alerts when nearing thresholds.
- If provider limits don’t exist:
- set limits on the payment card as a last safety layer, or
- disconnect payment after prepayment.
- Application-level controls
- Add code to analyze resource-intensive requests per user.
- If abuse is detected, restrict only that user/account, not everyone.
- Limit input data size (prevent prompt stuffing).
- Log agent calls to external tools.
- Prevent loops using max call counts (e.g., max 10 tool calls per session).
- Guardrails / prompt security
- Use guardrails such as:
- NVIDIA NeMo / NeMo framework
- Goal: detect whether prompts are intended to abuse or attack.
- Use guardrails such as:
- Bot-abuse protections
- Rate limiting (requests over time)
- Block/check suspicious IP reputation
- CAPTCHA-like challenges (e.g., Cloudflare)
- Stronger verification when needed:
- phone number requirements
- payment-card linking
- Provider-level limits
- Important constraint:
- Align controls with your risk model to reduce harm to legitimate users.
- Reminder:
- No single layer is perfect; layering increases attacker difficulty.
3) Prevent a bot from taking disastrous actions when it can send emails/quotes
- Recommended mechanism: human-in-the-loop
- Bot drafts emails/quotes
- Require a human to press “send”
- Rationale:
- Writing/creating offers takes the longest; manual review remains efficient.
- If later you trust automation:
- You may consider full auto-send, but the speaker advises caution.
- Extra safety option: add a second AI bot as reviewer
- The reviewer checks the first bot’s output before sending.
- Example checks:
- ensure the quote doesn’t contain extreme discounts (e.g., “200% discount”)
- stop sending outside working hours.
- Legal/communication fallback:
- Consult lawyers about adding a disclaimer/style note, e.g.:
- quotes prepared by AI may contain errors
- final price confirmed at order stage.
- Consult lawyers about adding a disclaimer/style note, e.g.:
4) Avoiding environment takeover: where developers must be careful
- Motivating examples:
- Recent account compromises of developers on popular projects (names recalled: “Lidl LM” and Axios).
- Lesson: even experienced developers make risky assumptions.
- Key workflow protections:
- Separate your private computer from the development environment
- If not, infection in a dependency can compromise:
- your data
- private files on disk
- If not, infection in a dependency can compromise:
- Use safer dev environments:
- a dedicated server
- connect via SSH tunnel
- or a dedicated virtual machine
- Docker caution
- In some configurations containers share the host kernel → higher risk.
- Least privilege everywhere
- Don’t install/run as administrator/root.
- Use dedicated users and restrict access.
- Don’t copy-paste untrusted commands blindly
- Especially commands that might be malicious (e.g., from a webpage) copied into
sudo. - Malicious payloads can be installed, including via clipboard manipulation.
- Especially commands that might be malicious (e.g., from a webpage) copied into
- Prefer package managers for installation
- Even packages can be compromised.
- Separate your private computer from the development environment
- Reduce supply-chain risk with:
- version pinning (force a known version)
- a cooldown period (ignore updates for a time)
- Example rationale: many malicious packages are detected early; waiting ~7 days can avoid worse outcomes.
- Block/restrict post-install scripts where possible.
- Use security options provided by your tools
- Ask an AI assistant about features; configure them (change config files/variables as needed).
- Hardening guides
- For each technology/tool, find and implement a “hardening guide”
- Search pattern:
<tool> hardening guide(e.g., npm + hardening guide, docker + hardening) - These settings are often what auditors check.
- Firewall on exit
- Don’t allow all processes to make outgoing connections freely.
- Benefit: even if malware steals data, it may be unable to exfiltrate.
- Example tools for interactive blocking/alerts:
- Little Snitch
- Alternatives named: Pod Portmaster, Class Wire, LuLu
5) Doing AI pentesting internally before release (and whether it’s enough)
- The speaker argues internal testing often isn’t sufficient due to cognitive biases:
- confirmation bias
- curse of knowledge
- creators may miss attack techniques they weren’t aware of
- Types of tools mentioned:
- static and dynamic security scanners
- red-teaming/self-awareness tools/models:
- Garak
- Pirate
- Deep Team
- Promptu
- Limitations:
- No tool guarantees 100% detection
- Tools may produce many false positives, wasting time on non-issues or non-exploitable findings.
- Recommended approach:
- Outsource security testing to professional experts
- Rationale: experts break systems professionally and objectively (no emotion).
- Market expectation: clients may require an independent professional team.
- Outsource security testing to professional experts
- Still, outsourcing doesn’t replace good engineering:
- Apply secure practices early (design stage, before coding)
- References suggested:
- GenI Security project (OASP foundation), especially AI SVS
- NIST AI RMF
- Final emphasis:
- Cybersecurity mistakes can cause major damage (not just reputation but also data leaks).
- Always run security checks before exposing the application to the internet.
Speakers / sources featured
Speaker(s)
- Unspecified primary speaker/host (appears to be the instructor): the presenter (name not provided)
Tools/platforms & external sources mentioned
- n8n (including self-hosted n8n)
- Make (Make.com)
- Unfcale (likely mis-transcribed; exact product unclear)
- HashiCorp Vault
- Google Secret Manager
- AES Secret Manager (as transcribed; exact product unclear)
- Volt (as transcribed)
- Cloudflare Tunnels
- Tailscale
- WireGuard
- Slack
- NVIDIA NeMo (as “Nvidia Nemo framework”)
- Cloudflare (rate limiting/CAPTCHA mention)
- Little Snitch
- Pod Portmaster / Class Wire / LuLu (as transcribed)
- Security testing tools/models:
- Garak
- Pirate
- Deep Team
- Promptu
- Standards / organizations / documents:
- OASP foundation (via GenI Security)
- AI SVS
- NIST AI RMF
- Libraries/providers named (as transcribed; possible transcription inaccuracies):
- “Lidl LM”
- Axios