Blog · Published · Updated
Prompt Injection in AI Agents: Controls for 2026
Short answer
Prompt injection is when an email, a document or a web page poses as an instruction for the model. In an agent with access to tools it can trigger an unwanted action. Several independent safeguards build the defence: least privilege, a separate reader for untrusted content and approval of high-impact actions. At Syntalith an agent costs from €3,000 net, and with 500 agent-handled cases a month at 8 minutes each it pays back in 1.7 months.
How we workDetails ↓
- 01Call30 minutes with an engineer, free
- 02PrototypeA working system on your process and data, free
- 03ImplementationTarget and fixed price in the contract; 50% at signing
- 04Acceptance50% once the agreed target is met
- net per agent for one process, build paid once
- from €3,000
- from signing the contract to acceptance
- 6–16 weeks
- free prototype on your cases
- €0
We belong to the OpenAI and Anthropic partner programmes and learn through their official courses. What this means ↓
Author: Syntalith · 3 min read
Statistics
Agent security: statistics 2026
1.7 mo.
payback of a €3,000 agent at 500 agent-handled cases a month: €2,327 of work today, €599 with checking and the model. Scenario.
48.8%
According to Eurostat (2025), 48.8% of EU companies that considered AI and did not use it named concerns about data protection and privacy.
Source: Eurostat, Use of artificial intelligence in enterprises (2025) ↗
49%
According to the European Commission (2026), 49% of EU SMEs and mid-caps name legal uncertainty as a barrier to using AI.
Source: European Commission, Flash Eurobarometer 578 (2026) ↗
40%
According to Gartner (2025), over 40% of agentic AI projects will be cancelled by the end of 2027 because of costs, unclear value or weak risk controls.
52.5%
According to Eurostat (2025), 52.5% of EU companies that considered AI and did not use it named a lack of clarity about the legal consequences.
Source: Eurostat, Use of artificial intelligence in enterprises (2025) ↗
Figures checked: 22 October 2026
What prompt injection is
A direct attack is an instruction the user types in: „ignore the previous instructions and…”. An indirect attack hides the instruction in content the agent reads along the way: in an email, a PDF, a web page, a ticket or a tool's output. OWASP ranks prompt injection as LLM01:2025 and separates it from a jailbreak: a jailbreak gets around the model's safeguards, and prompt injection changes how the application behaves. OWASP also describes a separate Top 10 for agentic applications for 2026.
In an agent the effects grow for three reasons. System instructions, the user's message, documents and tool results go into one context with no technical separation. A hidden instruction saved in memory or a RAG collection affects later sessions. The model runs files, email, APIs and MCP servers, so an instruction causes a change outside the chat window. OWASP names broad permissions as a risk of its own, Excessive Agency.
- 01Readera separate model reads untrusted content with no tools
- 02Contracta closed set of fields reaches the agent
- 03Permissionsonly the actions and data needed, no keys that bypass safeguards
- 04Policya rule outside the model before every call
- 05Approvala person approves irreversible actions
- An instruction hidden in an email and in a PDF attachment
- An instruction saved in a RAG collection
- An attempt to call a tool outside the scope
- An attempt to leak data in the answer
What a protected AI agent costs and when it pays back
- Today
- 500 agent-handled cases × 8 min = 66.7 h × €34.90 = €2,327
- With the system, a person checks the output in 25% of the time
- €582
- Model, 3 calls per case, Claude Sonnet 5.5
- €17
- Cost of one case
- €4.65 today, €1.20 with the system
- Saving
- €1,728 a month
Payback of €3,000
1.7 months
The volume, the time and the checking share are the scenario's assumptions. €34.90 an hour is the EU average hourly labour cost for 2025 according to Eurostat. Model: 4,000 input tokens and 500 output tokens per call at Anthropic's rates, converted at the ECB rate of 1.1205 USD per euro on 22 October 2026.
The count covers the work around a case: searching, reading, writing and passing it on. What stays is the time to check the output, 25% of today's work in the scenario. Before we quote, we measure both times on 20 to 100 cases your team has already handled, in a €0 test that takes 5 working days.
OWASP describes EchoLeak as an example of an instruction hidden in an email and used to leak data through a productivity system. Content the agent should only read then becomes the source of the attack, which is why the reader and the executor are separate components.
Can an AI agent delete the company database
Only if it was given the right to delete and there was no stop condition before an irreversible action. That is why the agent has the least access it needs, a separate technical account and no access to production. OWASP notes that there is no reliable method today to eliminate prompt injection completely, so acceptance covers attack tests, a permission review and a record of actions. A key that bypasses row-level security in the database should never reach the model. Content shown to users that comes from AI is also covered by the European Commission's guidelines on transparency.
The agent only reads and proposes
a separate reader and approval of every action
from €3,000
The agent performs routine actions
a policy layer and least privilege
from €3,000
An agent across several systems with sensitive data
attack tests at acceptance and monitoring
priced after the review
Where to start
Before a rollout we ask what the agent could do after a wrong decision, and we design it so the worst effect can be reversed. The criteria of an agent are described in the guide to AI agents. We build this kind of rollout as part of AI agents. For the free process review, a description of one case from recent weeks and the tools it went through is enough. You do not need to send confidential documents, and you get a written plan within 2 working days.
Worked example · Scenario
How we count the payback: an agent with safeguards
We run the same count for your process in the free process review.
€4,188
a month for searching the documentation for answers to 600 questions
Payback: 0.8 mo.
12 minutes to find an answer or ask a colleague and checking in 15% of the old time are our assumptions. €34.90 an hour is the EU average hourly labour cost for 2025 according to Eurostat, for companies with 10 or more employees. Model cost: Claude Sonnet 5.5, about 6,000 + 500 tokens per answer from the documents, at the ECB rate of 1.1205 USD per euro on 22 October 2026.
See the full calculation →
