Free AI prompts, planners & courses — grab them in the Freebies Library Get Freebies

Cloud Security for AI: Protecting Models, Data & APIs

Cloud AI adds new risks: stolen model weights, poisoned training data, leaked API keys. What to secure, which guardrails to use, and what it costs.
Cloud Security for AI: Protecting Models, Data & APIs

Quick Answer

AI workloads add six new assets to protect: training data, model weights, prompts, vector stores, GPU compute and inference APIs. Most 2025–2026 incidents were classic cloud mistakes — leaked keys, open buckets, unpatched servers — with AI-sized consequences. Good cloud hygiene, a guardrail in front of user-facing models, and a spending cap cover most of the risk. For the AI side of this picture, see what agentic AI is and how it works.

Why Do AI Workloads Change Cloud Security?

A traditional cloud app has three things to protect: code, a database and user traffic. An AI workload adds at least six more assets, and each one is a new target.

Training data is often the most valuable thing a company owns. If it leaks, the damage is the data itself plus the fact that it is now baked into a model that cannot be un-trained. Model weights are the trained model file — stealing them gives the attacker your product without your compute bill. Weights are large, hard to rotate like a password, and hard to detect when copied.

Prompts and logs routinely contain secrets, personal data and business plans, and they sit in logs, caches and vector stores long after the chat ends. Vector databases are the retrieval layer that lets a model answer from your documents; researchers have shown embeddings can leak the original content, and a poisoned vector store poisons every answer. GPU compute is the most expensive cloud resource most companies rent — an attacker who gets your quota can mine crypto or run inference on your bill. This is "denial of wallet": the bill, not the service, is the target. Inference APIs accept free-form text, so every request is untrusted input the model may act on.

The shared-responsibility line moves too. With a hosted model, the provider runs the model itself — but you still own the data you send, the prompts you write, the keys you use, the agents you connect, and the permissions you grant. The provider's "we do not train on your data" promise covers the model, not your misconfigured bucket.

The attack surface is the whole lifecycle — data collection, training, fine-tuning, deployment, inference, monitoring — and an attacker can enter at any stage. Most AI security incidents in 2025–2026 are still ordinary cloud incidents. Wiz's 2026 research: about 80% of cloud breaches in 2025 came from basic hygiene failures, not zero-days, and more than 85% of organisations now run AI services tightly coupled to sensitive data stores. AI does not create a new category of cloud vulnerability. It multiplies where classic errors can hurt, because the assets are worth more and the blast radius is bigger.

What Is the OWASP Top 10 for LLM Applications?

The OWASP Top 10 for LLM Applications 2025 (OWASP GenAI Security Project; 2025 edition published November 2024, still current as of 1 October 2026) is the single most useful checklist in this post. Here are all ten, in plain words.

Risk What it means in plain words
LLM01 Prompt Injection Malicious instructions hidden in text the model reads — a web page, a document, an email — that the model follows as if you wrote them.
LLM02 Sensitive Information Disclosure The model reveals data it should not: secrets in training data, private documents in a retrieval store, or details from other users' sessions.
LLM03 Supply Chain A poisoned library, a tampered model download, or a compromised plugin reaches you through a trusted source.
LLM04 Data and Model Poisoning Bad examples slipped into training data so the model learns the wrong thing, or a backdoor is planted in it.
LLM05 Improper Output Handling Model output is trusted blindly and passed into a database, a browser or a shell, where it becomes a real attack.
LLM06 Excessive Agency An agent gets more tools and permissions than its job needs, so one bad prompt becomes a real action: sending mail, deleting files, spending money.
LLM07 System Prompt Leakage An attacker tricks the model into revealing its hidden system instructions, including secrets and rules meant to stay private.
LLM08 Vector and Embedding Weaknesses New in 2025. Poisoned documents in a retrieval store, text reconstructed from embeddings, or a vector store answering from documents the user should not see.
LLM09 Misinformation The model confidently produces false information that users act on, because it sounds right.
LLM10 Unbounded Consumption No quotas or spending caps, so an attacker — or a runaway agent — burns through your budget. Denial of wallet.

LLM07 and LLM08 are new versus the 2023 list. OWASP also released a Top 10 for Agentic Applications in December 2025. The frameworks behind the list: MITRE ATLAS (v2026.09) holds 16 tactics, 120 techniques and 73 case studies. NIST's AI Risk Management Framework gives four functions — Govern, Map, Measure, Manage. Google's SAIF is six principles, not a product. ISO 42001 is certifiable; ISO 27090 (14 October 2026) is not. A certificate proves AI is governed, not that any model resists attack.

Prompt injection has no complete fix. This is the single most important technical fact in this post. A language model cannot reliably distinguish instructions from data, because both arrive as text. Guardrails reduce the risk — screening, sandboxing, human approval for consequential actions — but no vendor claims to eliminate it. Anyone who does is selling something.

What Actually Went Wrong in 2025–2026?

Every incident below is a real, sourced event. The pattern is the story: familiar cloud failures with AI-flavoured consequences.

Incident What happened Lesson
302.ai — October 2025 An AI platform left an Azure bucket unsecured: 17.6 million files, 22.72 TB, including 144,000+ chat logs with prompts and personal data. Discovered 2 August, flagged 11 August, not observed closed until 8 September. Cybernews found no evidence anyone had copied the data. A bucket with no authentication is a public website, and AI companies hold chat logs in buckets.
Langflow CVE-2025-3248 and JadePuffer — 2025–2026 Unauthenticated remote code execution (CVSS 9.8) in the open-source AI agent builder, via code that ran exec() before the authentication check. Fixed in Langflow 1.3.0 (31 March 2025); added to CISA's Known Exploited Vulnerabilities list 5 May 2025. The Flodrix botnet exploited it in mid-2025, and VulnCheck counted 15,000+ successful exploitation attempts in April–June 2026 alone. Sysdig's JadePuffer disclosure (1 July 2026) called it the first publicly documented fully agentic ransomware: an autonomous agent that entered through this flaw, fixed its own failed login in 31 seconds, then encrypted and extorted the victim. AI tools are internet-facing servers. Patch them like servers — an unpatched one is now exploitable by another AI, at machine speed.
ShadowRay 2.0 — November 2025 An active global botnet attacked Ray, the distributed-compute framework behind much AI training, through CVE-2023-48022 (CVSS 9.8). The maintainers deliberately did not patch it — Ray is "not intended for use outside a strictly-controlled network environment". The first campaign stole cloud credentials, database passwords, private SSH keys and proprietary LLM weights. The 2025 campaign found 230,000+ Ray servers exposed and used LLM-generated payloads. A framework that assumes a trusted network is a loaded gun on the public internet, and AI training clusters are high-value targets.
Ultralytics PyPI hijack — December 2024 Four releases of the popular YOLO computer-vision library shipped a Monero miner. Two went out through the project's own release pipeline after GitHub Actions cache poisoning; two more through a leftover PyPI token. YOLO has about 60 million total PyPI downloads. A trusted publisher's own release pipeline was the attack path, and the malicious code ran inside every training job that installed the package.

Two more 2025–2026 supply-chain cases. The Fireworks AI breach (October–November 2026) spread to Cursor: a key Cursor had issued to Fireworks was exposed, a read-only bucket of coding-agent traces was copied, under 0.01% of Cursor customers had API credentials exposed, and tampered Claude Code extension copies appeared on Cursor's marketplace cache on 29 October. And the Model Context Protocol has no built-in way to detect when a server changes its definition after the user consents: Invariant Labs showed a poisoned tool stealing SSH keys, the postmark-mcp package shipped 15 clean versions then added a hidden BCC to an attacker's address, and GitGuardian found 24,008 secrets in public MCP configuration files, 2,117 still valid.

Can a Model Be Stolen or Poisoned?

Model extraction is demonstrated research, not a live theft. Carlini and colleagues (ICML 2024, best paper) recovered the full embedding projection layer of OpenAI's ada and babbage models for under $20 in API queries, and estimated the full gpt-3.5-turbo projection matrix at under $2,000. Nasr and colleagues extracted 10,000+ training examples from ChatGPT for about $200. An API that answers unlimited queries is a model-theft channel; rate limits and query monitoring are the defence.

DeepSeek distillation is an allegation, not proven theft. Bloomberg reported in January 2025 that Microsoft security researchers had observed individuals believed to be linked to DeepSeek exfiltrating a large amount of data via OpenAI's API, and OpenAI said it was reviewing indications that DeepSeek may have inappropriately distilled its models. No claim has been tested in court, and as of October 2026 no American developer had sued. A joint FBI, NSA and CISA advisory in September 2026 named six China-based companies. An allegation with a paper trail — never a proven theft.

Data poisoning is proven in research, not yet in a large production incident. Nightshade, a defensive tool from the University of Chicago, showed fewer than 100 poisoned images could corrupt one Stable Diffusion XL prompt — and a 2025 study built LightShed, which detected the poison with 99.98% accuracy in its experiments. PoisonGPT altered one fact in a model and uploaded it to Hugging Face under a typo-squatted name; it passed standard benchmarks while reliably lying about the targeted fact. Anthropic's Sleeper Agents research showed a backdoor can survive safety training, and a 2025 study found roughly 250 poisoned documents were enough to install a backdoor in models up to 13 billion parameters. No large-scale production poisoning incident is confirmed as of November 2026 — the risk is real because the mechanism is proven and the attack is cheap.

The autonomous-breach case is one case, and one company's inference. Wiz reported on 19 October 2026 a single multi-cloud case in which an AI agent ran an entire cloud breach end to end, entering through a Java remote-code-execution flaw in an exposed Kafka UI. This is Wiz's inference, not independently corroborated — do not write "AI agents are now hacking clouds" as established fact. Related Wiz reporting counted 9 AI developer-platform incidents and about 2.5 million Amazon Bedrock calls linked to a single vendor breach (one headline says 25 million; the body says 2.5 million), with attribution of all 9 incidents to one breach not established.

Leaked Keys and Shadow AI: The Problems That Actually Hit Small Users

Forget exotic attacks for a moment. The two risks most readers actually face are a leaked API key and shadow AI.

Leaked API keys. GitGuardian's State of Secrets Sprawl 2026 (17 March 2026) counted 28.65 million new hardcoded secrets in public GitHub commits in 2025, up 34% year on year — the largest annual jump in the report's history. AI-service secrets reached 1,275,105, up 81%; 12 of the 15 fastest-growing leaked secret types were AI services. The report counted 113,000 DeepSeek API keys in 2025 and 130,000+ Hugging Face tokens per year. Claude Code-assisted commits leaked at 3.2% versus a 1.5% baseline, and 64% of secrets confirmed valid in 2022 were still valid in January 2026. Private repositories were about 6 times more likely to contain hardcoded secrets than public ones (32.2% versus 5.6%). A leaked AI key is not just a billing problem — it is compute, data and model access in one string.

Shadow AI. IBM's Cost of a Data Breach Report 2025 (Ponemon, 600 organisations) put the global average breach cost at $4.44 million, and the US average at $10.22 million, an all-time high. Shadow AI — employees using AI tools the company never approved — was present in 20% of breaches. Organisations with high shadow AI averaged $4.74 million per breach versus $4.07 million with low or none. Shadow-AI breaches compromised personal data in 65% of cases (versus 53% global) and intellectual property in 40% (versus 33%), and 63% of breached organisations had no AI governance policy. Two further figures come from secondary analysis, not IBM itself: a National Law Review analysis found 97% of organisations with an AI-related incident had no proper AI access controls, and a secondary source says shadow-AI breaches took 10 extra days to identify and contain. Plainly, shadow AI is the cloud-AI version of shadow IT — and it is measurably more expensive when it goes wrong.

Wiz found in September 2026 that about 1 in 10 of roughly 3,000 internet-facing LiteLLM gateways — the open-source proxy many small teams use — accepted a default master key or no authentication at all, and a chain including an MCP authentication bypass led to full cloud compromise. And Wiz's GhostApproval finding (July 2026) showed a symlink trick letting a malicious repository make six AI coding assistants write outside their workspaces, including Claude Code, Cursor and Amazon Q Developer. An AI coding assistant with file-write access is a privileged user, and a repository is untrusted input.

Who Secures What in the Cloud?

The model every cloud uses is the same: the provider secures the infrastructure, and the customer secures what they put on it and how they configure it. AI adds a middle layer — the model itself — that sits mostly with the provider. But the data, prompts, keys, agents and permissions are always the customer's.

AWS (Bedrock / SageMaker) Azure (Azure OpenAI / Foundry) Google Cloud (Vertex AI / Gemini Enterprise)
Provider secures The infrastructure, the physical data centres, and the managed service itself. The infrastructure and the Foundry platform. The infrastructure and the Vertex AI platform.
You secure Your data, IAM roles and policies, network configuration, encryption keys, and the prompts and outputs you send. Your data, identities, network, and the applications you build. Your data, IAM, network, and applications.
Training on your data Customer inputs and outputs are NOT used to train base models and are NOT shared with model providers. Prompts and completions are NOT used to train foundation models without permission and are NOT available to OpenAI. EU DataZone processing is available for EU data residency. Service terms restrict training or fine-tuning on customer data without prior permission or instruction.
Retention caveat Abuse-detection retention: classifier-flagged traffic kept up to 30 days for most OpenAI models; ALL traffic kept up to 30 days for Claude Fable 5 and 5.1, and may be human-reviewed by AWS. Abuse monitoring inspects flagged content unless the customer is approved for modified abuse monitoring. Gemini models cache data in memory and, for supported models, durable storage for up to 24 hours by default (can be turned off). Prompts may be logged for abuse monitoring.
Key security features SageMaker VPC-only mode, EnableNetworkIsolation, KMS encryption of S3 data, EBS volumes and model artifacts, VPC endpoints, least-privilege execution roles. Inter-container traffic encryption is opt-in, not on by default. The default Foundry guardrail blocks hate, sexual, violence and self-harm content at medium severity and detects direct prompt attacks. No separate published charge for the default guardrail. Data-residency and region controls apply as for other GCP services.

The rule of thumb: wherever a company signs a contract (API tier, business tier), training is off by default. Wherever an individual opens a free account, it is on. That is the bridge to the consumer side — our AI data privacy guide covers what happens when a person types into a chat tool, including training opt-outs and retention windows. This post is the cloud side. For enterprises, Wiz offers AI security posture management (quote-only pricing) — treat it as the enterprise option, not the small-business default.

Which Guardrails and Detection Tools Exist, and What Do They Cost?

All prices below were verified against the provider's own pricing page in October–November 2026. Prices move, so check the live page. A text unit or record is up to 1,000 characters or code points unless stated otherwise.

Service What it does Price (as of November 2026)
Bedrock Guardrails (AWS) Screens prompts and responses: content filters, denied topics, sensitive-information filters, contextual grounding checks, Automated Reasoning checks, word filters. Content filters $0.15, denied topics $0.15, sensitive-information filters $0.10, contextual grounding $0.10, Automated Reasoning $0.17 per 1,000 text units per policy; word filters and regex filters FREE; images $0.00075 each. No Guardrails-specific free tier. It does NOT evaluate model reasoning content blocks — a real gap.
InvokeGuardrailChecks API (AWS, June 2026) Calls Bedrock's guardrail checks without running the model through Bedrock — useful for screening before sending to another model. Content $0.07, prompt attack $0.08, sensitive information $0.10 per 1,000 text units.
Model Armor (Google Cloud) Screens prompts and responses for prompt injection, harmful content, sensitive data and bad links or URLs. Works with models outside GCP. Standalone: first 2 million tokens a month FREE, then $0.10 per million tokens. Included in Security Command Center Premium and Enterprise tiers: 3 billion tokens a month included on subscription, then $0.10 per additional 1 million tokens. Limits: it judges each request alone (no conversation memory) and does not decode Base64.
Azure AI Content Safety + Prompt Shields (Microsoft Foundry) Content Safety screens text and images for harmful content. Prompt Shields detect direct user-prompt attacks AND document attacks (up to 5 documents). Free tier F0: 5,000 text records + 5,000 images a month (5 requests a second). Standard S0: $0.375 per 1,000 text records, $0.75 per 1,000 images; Prompt Shields billed on the same text-record meter; commitment tier $338 a month for 1 million text records (overage $0.338 per 1,000). Document attack detection is OFF by default — turn it on. Microsoft warns a clean Prompt Shields result "doesn't establish that the request is benign".

One third-party estimate is worth quoting because the gap is large. ecorp-it (October 2026) put screening 10 million agent requests a month (about 2,700 characters each) at roughly $675 with Model Armor, $4,700–$10,000 with Bedrock Guardrails, and $18,750 with Azure Content Safety + Prompt Shields. A third-party estimate, not a vendor price.

Service What it does Price (as of November 2026)
Amazon GuardDuty Threat detection: monitors CloudTrail, VPC Flow Logs, DNS logs, S3, EKS audit logs, runtime activity, malware scanning. Pay-as-you-go, no flat fee; 30-day free trial per Region and plan. Key meters: CloudTrail $4.00 per 1 million events; VPC Flow Logs + DNS from $1.00 per GB; S3 Protection $0.80 per 1 million events; EKS audit logs $1.60 per 1 million; Runtime Monitoring $1.50 per vCPU; Malware Protection $0.03 per GB (EC2) and $0.09 per GB + $0.215 per 1,000 objects (S3). A small account typically runs $2–$100 a month (third-party estimate).
AWS Security Hub Central security findings and compliance checks across AWS accounts. Pricing changed in 2026: now a per-resource "Essentials plan" that folds in Inspector scanning and CSPM checks. The old per-check price is outdated and is not quoted here. No current per-check price could be verified. GuardDuty threat detection is a separate add-on.
Microsoft Defender for Cloud — AI workloads Detects prompt injection, jailbreak attempts and data leakage via Prompt Shields; discovers the AI bill of materials across Foundry, Azure OpenAI, Vertex AI and Bedrock. Plan price NOT verified — no price is quoted here. The AI-workloads plan was in preview or limited preview as of November 2026. A preview AI Agent Protection feature auto-discovers agents, monitors audit logs, and can block an agent's response in real time on prompt-injection detection.
Google Security Command Center (SCC) Cloud security posture and threat detection for GCP. Model Armor is included in Premium and Enterprise tiers. Standard tier free. Premium: minimum annual cost $15,000 per a UK G-Cloud pricing document (January 2026) — reported, not from Google's public page. Enterprise: fixed-price subscription, no public price.

Secrets are the cheapest security money can buy. AWS Secrets Manager: $0.40 per secret a month + $0.05 per 10,000 API calls; rotation adds $0.05. Parameter Store standard parameters are FREE. Azure Key Vault: $0.03 per 10,000 operations; vault fees ($1 a month Standard, $6 a month Premium HSM) come from a third-party source. Google Secret Manager: first 6 active secret versions FREE, then $0.06 each a month; first 10,000 access operations free, then $0.03 per 10,000. One third-party calculation puts 100 secrets + 100,000 calls a month at about AWS $45, Azure about $3, GCP about $9 — flag it as an estimate. The point: a secrets manager costs less than one leaked key.

What Should a Small Business Do First?

Start with the free steps, in this order.

1. Move every API key out of code and repositories into environment variables (free) and rotate any key that was ever committed. A leaked key is compute, data and model access in one string.

2. Turn on multi-factor authentication on every cloud and AI account (free).

3. Turn on the free tiers: GuardDuty 30-day trial, Model Armor 2 million free tokens a month, Azure Content Safety F0 5,000 records a month, GCP SCC Standard (free).

4. Set billing alerts and a spending cap on every AI service (free). The LLMjacking case — attackers linking their own Google Cloud projects to victims' billing accounts and running Gemini usage the victims could not see, with charges in the tens to hundreds of thousands of dollars within 1–2 days — is the reason.

5. Make every storage bucket private and turn on Block Public Access (free).

6. Turn off training on your data where the toggle exists, and use business or API tiers where training is off by default (free to change, may cost more per call — say so honestly).

7. Write down which AI tools your team uses (the shadow-AI inventory — free, and the single cheapest governance step there is). For tooling that helps with this, see our guide to AI governance tools.

Beyond the free list, the same good cloud practices apply, with AI-specific edges. Keys first: never hardcode a key in code, a notebook or a repository — use a secrets manager or environment variables that are never committed, and give each key the smallest scope the provider allows. Least-privilege IAM: every identity — person, training job, endpoint, agent — gets only the permissions it needs. A training job does not need to read your production database. Never let a notebook or an agent assume an admin role.

Private endpoints and VPC: put training jobs, endpoints and vector stores in a private VPC with no public internet access, and reach cloud services through VPC endpoints (PrivateLink). SageMaker has a VPC-only mode and EnableNetworkIsolation — use them. Encryption: encrypt data at rest with customer-managed keys and in transit everywhere. Encryption protects data if storage is stolen, but not if the attacker has a valid key or login. Necessary, not sufficient.

Logging and monitoring: turn on CloudTrail, Azure activity logs or Cloud Audit Logs, and keep them in a separate, locked-down account or bucket — an attacker who can delete the logs can hide the breach. Turn on GuardDuty, Defender or SCC even on the free tier. Treat prompt logs as sensitive data: they contain the same secrets the prompts did. Cost guardrails are security guardrails: per-key quotas, per-day spending caps, rate limits, and alerts at 50%, 80% and 100% of budget.

Screen inputs and outputs: put a guardrail in front of any model that faces users or untrusted documents. The cheapest options cost nothing to start. Remember the limits: no guardrail catches everything, document-attack detection in Azure is off by default, and Bedrock Guardrails do not evaluate reasoning blocks. Agents get boundaries, not freedom: an agent that can send email, delete files or spend money needs a whitelist of allowed tools, a budget cap, a maximum number of steps, human approval for anything that sends, pays, publishes or deletes, and a log of every action. Start at "suggests, human acts" and only move to "acts within limits" when the action is bounded, reversible and checkable — the autonomy spectrum in our AI agents and automation guide. Supply chain: pin library versions, never install "latest" in production, check a model hub download like a binary, never load a pickle-serialised model from an untrusted source, and re-review every MCP server when it changes, because the protocol will not warn you.

What Does the Law Require?

EU AI Act (Regulation (EU) 2024/1689, in force 1 August 2024). The law sorts AI by risk, not by technology. The relevant duty is Article 15: high-risk AI systems must achieve "an appropriate level of accuracy, robustness, and cybersecurity" throughout their lifecycle and stay resilient against named AI-specific attacks — data poisoning, model poisoning, adversarial examples and confidentiality attacks. After the Digital Omnibus (Regulation (EU) 2026/1744, in force 27 July 2026), Annex III high-risk duties apply from 2 December 2027, and Annex I from 2 August 2028. General-purpose AI duties applied from August 2025. Serious incidents must be reported within 15 days (10 days for a death, 2 days for widespread incidents). Fines reach €35 million or 7% of worldwide turnover for prohibited practices, €15 million or 3%, and €7.5 million or 1%. Plainly: if your AI system decides who gets a job, a loan or a university place in the EU, the law now requires you to prove it resists poisoning and extraction — deadline December 2027. Most small businesses building ordinary tools are NOT high-risk.

GDPR. The AI Act is about AI safety; GDPR is about personal data, and it still applies. Article 32 requires appropriate technical and organisational measures — encryption, access control, testing — for any personal data in an AI workload, and Article 35 requires a Data Protection Impact Assessment where processing is likely high-risk. A training dataset with personal data, a vector store built from customer documents, and a prompt log full of names are all GDPR-relevant. Pakistan has no comprehensive data-protection law (the Personal Data Protection Bill has been a draft since 2018; PECA 2016 section 16, up to 3 years or Rs 5 million) — one line here, because the Data Privacy post covers it in full.

SOC 2 and ISO 27001 — what enterprise customers ask for. SOC 2 is an attestation (a report by an auditor), not a certificate. ISO 27001 is a certifiable information-security management system. Both are what enterprise customers ask an AI vendor for before they buy. For a small business, the practical point is the reverse: when YOU buy an AI service, ask the vendor for their SOC 2 report or ISO 27001 certificate. It is the fastest way to check whether they take security seriously. ISO 42001 (certifiable) and ISO 27090 (NOT certifiable, 14 October 2026) are the AI-specific additions. A certificate proves a system is governed, not that any model resists attack.

Key Takeaways

  • AI adds six new assets to protect, and most 2025–2026 incidents were classic cloud mistakes with AI-sized consequences.
  • Prompt injection has no complete fix. A model cannot reliably tell instructions from data, because both arrive as text. Guardrails reduce the risk; they do not eliminate it.
  • Leaked keys and shadow AI are the risks most readers actually face: 28.65 million new hardcoded secrets in public GitHub commits in 2025, and shadow AI present in 20% of breaches.
  • The provider secures the infrastructure and the model; you secure your data, keys, prompts, agents and permissions.
  • Guardrails are cheap to start: Model Armor's first 2 million tokens a month are free, Azure Content Safety's free tier covers 5,000 records a month, and a secrets manager costs a few dollars a month.
  • The EU AI Act's cybersecurity duty for high-risk systems applies from December 2027. Most small businesses are not high-risk — but the customers you sell to will still ask for your SOC 2 report or ISO 27001 certificate.

Frequently Asked Questions

Does AWS train on my Bedrock prompts?

No. AWS says Bedrock customer inputs and outputs are not used to train base models and are not shared with model providers. Caveat: abuse-detection retention applies — classifier-flagged traffic is kept up to 30 days for most OpenAI models, and ALL traffic for Claude Fable 5 and Fable 5.1, which may be human-reviewed by AWS.

Can prompt injection be fully prevented?

No. A language model cannot reliably distinguish instructions from data, because both arrive as text. Guardrails, sandboxing and human approval for consequential actions reduce the risk, but no vendor claims to eliminate it. Anyone who does is selling something.

How much does a guardrail cost?

You can start free: Model Armor's first 2 million tokens a month are free, and Azure Content Safety's free tier covers 5,000 text records a month. Paid screening runs roughly $0.07–$0.17 per 1,000 text units on AWS, $0.10 per million tokens on Google Cloud after the free tier, and $0.375 per 1,000 text records on Azure. Prices are as of November 2026 — check the live pricing page.

What is shadow AI?

Employees using AI tools the company never approved — signing up with a work email and pasting company data into them. IBM's 2025 breach report found shadow AI present in 20% of breaches, with those breaches costing about $670,000 more on average than breaches at organisations with low or no shadow AI.

Is my API key safe in a private GitHub repo?

No. GitGuardian's 2026 report found private repositories were about 6 times more likely to contain hardcoded secrets than public ones (32.2% versus 5.6%). Never hardcode a key anywhere. Use a secrets manager or environment variables, and rotate any key that was ever committed.

Do I need to comply with the EU AI Act?

Only if your AI system is high-risk under the Act — recruitment, credit scoring, education or law enforcement in the EU, for example. Most small businesses building ordinary tools are not high-risk. If you are, the cybersecurity duty (Article 15) applies from 2 December 2027.

Want the consumer side of AI privacy? Read our AI data privacy guide, and for the agents themselves, see our AI agents and automation guide. Grab the freebies in our Freebies Library.

About the Author

Scholarships, free courses & AI tools for students.
LinkedIn

Post a Comment

Cookie Consent
We serve cookies on this site to analyze traffic, remember your preferences, and optimize your experience.