AI safety and responsible deployment

Why safety is part of the job

An AI assistant inside a bank, hospital or government office can affect real people's money, health and rights. "The model said so" is not an excuse. Whoever builds and deploys the system is responsible for how it behaves. Safety is designed in from the start, like security in software.

The main risks

• Hallucination: fluent but false answers (module 10).
• Prompt injection: instructions hidden in user input, documents, emails or web pages that try to override your rules. Indirect injection arrives through retrieved content the user never typed.
• Data leakage: the model reveals personal data, confidential documents or its system prompt.
• Jailbreaking: users trying to trick the model into breaking its rules.
• Harmful, biased or toxic output.
• Over-powered agents: a model that can send money, delete records or email customers without checks.

Defence in depth

1. Least privilege: give the model and its tools only the permissions they truly need. A support bot can look up an order; it should not be able to issue refunds on its own.
2. Human approval for risky or irreversible actions (payments, deletions, messages to customers).
3. Separate instructions from data: mark retrieved or user-supplied text as untrusted content, and never let it change the system's rules.
4. Input and output filters: block or flag personal data, toxic content and off-topic requests; validate output formats.
5. Mask personal data before it is sent to a model, and do not log more than you need.
6. Grounding and citations so answers can be checked.
7. Logging and monitoring, with a clear incident process to switch features off quickly.

import re

EMAIL = r"[A-Za-z0-9._%+-]+@[A-Za-z0-9.-]+\.[A-Za-z]{2,}"
MOBILE = r"(?:\+91[ -]?)?(?<!\d)[6-9]\d{9}(?!\d)"

def mask(text):
    return re.sub(MOBILE, "[PHONE]", re.sub(EMAIL, "[EMAIL]", text))

print(mask("Call 9876543210 or mail asha@example.com"))

Red-teaming

Red-teaming means deliberately trying to make your own system fail before attackers or users do: injection attempts, requests for other customers' data, harmful requests, tricky languages and formats. Add every failure you find to the evaluation set (topic 3) so it stays fixed. The OWASP Top 10 for LLM Applications is a useful checklist of common risks.

Transparency and law

Tell users when they are talking to an AI, and give a way to reach a human. Label AI-generated content where it could mislead. In India, personal data processing must follow the Digital Personal Data Protection Act, 2023 (module 12); sector regulators such as RBI may add their own rules. Check the current rules for your sector.

Common mistakes

• Relying on the system prompt alone ("never reveal secrets") as the only defence.
• Giving an agent broad API keys "to keep things simple".
• Logging full prompts with customer personal data forever.
• Testing safety once before launch and never again.