The scariest sentence in AI security is not typed by an attacker in a hoodie. It is typed by a regulator, calmly, in guidance prose: you cannot fix this inside the model. Somewhere, a prompt engineer just spilled coffee on a laminated list of magic words. That is the useful bit in the Australian Signals Directorate’s new framing of prompt injection for agentic AI. ASD is not treating prompt injection like a flaky unit test, one more bug to squash before demo day. It is pointing at the shape of the system itself, which is much less comforting and much more actionable.

ASD moves the fix outside the model

According to iTnews, ASD has issued new guidance for enterprises saying a central security weakness in agentic AI cannot be resolved inside the model. The agency’s recommendation, as reported by iTnews, is that mitigation belongs in the software layer wrapped around models. InfoSec Today summarized the same core point: the weakness is not something model internals can permanently eliminate, so builders need controls around the model rather than heroic prompt incantations.

That matters because the risk rises when LLMs are connected to tools, data stores, and workflow actions. A chatbot that writes a bad haiku is embarrassing. An agent that reads hostile instructions from a document and then uses an approved tool with excessive privileges is a Roomba with a corporate credit card.

Why smarter prompting hits a wall

Alexander Arvidsson’s ArcticDBA analysis explains the architectural snag in plain terms: system prompts, user messages, and retrieved documents all flow through the same context window as tokens. In other words, the model receives instructions and potentially hostile content through the same general channel. That is less like having a locked filing cabinet and more like writing the password on a sticky note, then asking the sticky note to behave ethically.

Nilesh Barla at Adaline makes the complementary point that prompt injection is not fixed by better prompts because the attack surface lives in the tool layer. Arun Baby’s analysis similarly argues that filters are not enough, because LLMs process trusted instructions and untrusted data through the same token stream. The defenses Arun Baby identifies as stronger are architectural ones, including privilege separation, type directed isolation, and spotlighting.

This is where the conversation gets blessedly less mystical. The practical lesson is not stop using agents. It is stop pretending a longer system prompt is a seatbelt, an airbag, and a traffic law.

Research is converging on containment

Ciphers Security reported that a paper submitted to arXiv on May 17, 2026 by Sahar Abdelnabi and Eugene Bagdasarian examines why AI agents may always be vulnerable to prompt injections. The arXiv paper, AI Agents May Always Fall for Prompt Injections, says existing data instruction separation defenses can miss contextual manipulation and can also degrade contextually appropriate behavior. That is the miserable tradeoff: tighten the rules too much and useful work breaks, loosen them and an adversary can make a bad flow look legitimate.

This is not a brand new gremlin that crawled out of the GPU rack yesterday. The older arXiv paper Prompt Injection: Parameterization of Fixed Inputs shows the problem has had a formal research trail for years. If a vendor pitch treats prompt injection as a newly discovered inconvenience solved by proprietary fairy dust, check the bibliography before you check the invoice.

The 2026 arXiv paper reframes the problem through Contextual Integrity, a privacy theory about whether information flows comply with contextual norms. That framing is helpful because agents increasingly make decisions based on context, not just static strings. The hard part is not spotting the phrase ignore previous instructions, it is recognizing when a seemingly reasonable request smuggles in the wrong authority, purpose, or flow.

What builders should do now

Taken together, ASD’s guidance via iTnews and the research literature point toward containment. Give agents the least tool access they need, separate trusted instructions from untrusted content as much as the architecture allows, require human review for high impact actions, and monitor tool use like it belongs to a very eager junior employee. Not because the junior employee is evil, but because it can be convinced the photocopier is a database migration strategy.

For product teams, the immediate question is not whether an LLM can be made perfectly obedient. It is what happens when it is not. Design systems so a bad instruction hits permission boundaries, logging, review gates, and scoped tools before it touches money, customer data, or production workflows.

For readers building with agents, ASD’s warning is a nudge away from prompt superstition and toward security engineering. The model is not the castle. It is the very persuasive intern holding the keys.

Sources - ASD says prompt injection in AI cannot be fixed - iTnews

Sources