By Adrian Culley, Director of Solutions Engineering, EMEA (Field CTO), SafeBreach
Over the last two years, enterprises have connected language models to email, code repositories, ticketing systems, customer data and cloud consoles, often faster than security teams could review them. In 2027, the consequences of that speed will arrive. Prompt injection, over-privileged agents and poisoned context, long confined to research papers and conference talks, will appear in mainstream incident reports. Four shifts will define the year.
Prediction 1: Agents become the most privileged identities nobody is governing.
An AI agent with tool access is effectively a non-human identity that reads untrusted content, makes decisions and acts with real credentials. Most organisations cannot yet say how many agents they run, what each can touch, or who approved it. I expect 2027 to bring the first widely reported breaches where the root cause is an agent granted far more authority than its task required. The OWASP Top 10 for LLM Applications names this risk as excessive agency, and it will move from checklist item to board-level concern. The fix is familiar: least privilege, scoped and short-lived credentials, human approval for high-impact actions, and a maintained inventory.
Prediction 2: The protocol layer becomes the new perimeter.
Standards such as the Model Context Protocol make it easy to plug models into tools and data, and just as easy to create unvetted trust boundaries. Shadow MCP servers, unsigned tool definitions, weak parameter validation and unsandboxed execution will be to 2027 what exposed S3 buckets were to 2017: an unglamorous, widespread and avoidable source of breaches. Government guidance on securing these deployments is already emerging, and I expect auditors to begin asking for MCP inventories and logging as routine evidence.
Prediction 3: Indirect prompt injection goes industrial.
Attackers have learned that they do not need to talk to your model; they only need to get text in front of it. Malicious instructions hidden in web pages, documents, calendar invites, support tickets and code comments will be used to steer agents into leaking data or taking actions. Because language models cannot reliably separate instructions from data, no filter will fully solve this. Defence will therefore rest on architecture: isolating untrusted content, constraining what a compromised agent can do, validating outputs before they reach downstream systems, and assuming that some injections will succeed.
Prediction 4: Attackers scale with AI, and defenders must prove their controls hold.
Offensive use of AI will keep compressing the time from vulnerability disclosure to working exploit, and make phishing, reconnaissance and malware adaptation cheaper. Defenders will respond with AI in the SOC, but a new risk follows: models that triage alerts or write detections are themselves targets for poisoning and evasion. The organisations that cope best will test their AI-enabled defences the way they test everything else: red-teaming models and agents, simulating realistic attack techniques, and checking that guardrails and detections actually fire. Assurance will come from evidence, not vendor claims.
Prediction 5: Regulation catches up, and demands proof.
The EU AI Act, NIS2, DORA and national frameworks are converging on a common expectation: demonstrate that AI systems are governed, monitored and resilient. In 2027, I expect supervisors to ask pointed questions about AI inventories, human oversight, incident reporting and testing cadence. Frameworks such as NIST AI RMF, ISO 42001 and MITRE ATLAS will become the shared vocabulary between security teams, auditors and boards.
What this means for practitioners.
Start with visibility: list every model, agent and MCP connection in use, including the ones teams adopted without asking. Apply least privilege to agents exactly as you would to service accounts. Treat all model inputs and outputs as untrusted. Log agent actions in a form your SIEM can use. And build a regular cycle of adversarial testing for AI systems, mapped to OWASP and ATLAS, so that you can show your controls work rather than assume they do.
AI will not make security obsolete, nor will it doom it. It will simply remove the room for guesswork. In 2027, the organisations that treat AI systems as part of the attack surface, and test them accordingly, will be the ones that benefit from the technology without becoming its cautionary tale.
##
ABOUT THE AUTHOR

Adrian Culley is Director of Solutions Engineering, EMEA at SafeBreach, where he acts as Field CTO. He is a core team contributor to the OWASP Top 10 for LLM Applications and has a long-standing focus on offensive security, adversary emulation and threat intelligence.






