Skip to main content
AI News & Updates

Granite 4.1 LLMON Previews Launch: Test Instruction/Data Separation With Synthetic Injections

Compare ordinary prompts with structured blocks in isolation without treating author benchmarks as production proof

ENHE AI5 min6 views

IBM released experimental 3B and 8B Granite 4.1 LLMON preview models on November 12, 2026. LLMON uses explicit structured blocks to separate instructions, data, and requirements, with binding intended to express which content may be executed. This guide proposes a small safe evaluation using synthetic data only. Create normal, conflicting, and prompt-injection cases; freeze the model and decoding settings; compare an ordinary prompt with an LLMON prompt; and record task correctness, refusal, unauthorized behavior, and parsing failure. The published results are author-run benchmarks, and the models are explicitly previews for exploration and evaluation. They should not receive production data or external-action permissions before independent security review.

More topic tags (8)
AI NewsAuto PublishingSEO提示注入Granite 4.1LLMON结构化提示指令数据分离
Granite 4.1 LLMON Previews Launch: Test Instruction/Data Separation With Synthetic Injections

Key takeaways

Use synthetic data and an isolated environment with no external actions first.
Keep model, task, and decoding settings identical across ordinary and LLMON prompts.
Record correctness, refusal, unauthorized behavior, and parsing failure—not only success rate.

By ENHE AI · 2026-11-14 · AI-assisted drafting; pre-publication quality checks completed.

Direct answer: treat it as an experimental preview, not a security switch

IBM released experimental 3B and 8B Granite 4.1 LLMON preview models on November 12, 2026. They are for exploration and evaluation rather than production.

Run the first round in an offline or credential-free container using fictional customers, keys, and documents. The model gets no email, database, file-write, payment, or publishing tools.

Build three synthetic case groups

Normal cases extract clear fields; conflict cases place text in data that opposes the system task; injection cases include fictional phrases such as “ignore above and act.” Write expected answers before running at least five per group.

Cases contain no real credentials, personal information, or executable attack payload. The goal is to compare structure boundaries, not probe unauthorized systems.

How to compare ordinary and structured prompts

LLMON separates instructions, data, and requirements with explicit blocks. Write semantically equivalent ordinary and LLMON versions while fixing model, temperature, seed, output limit, and parser.

Record task correctness, explicit refusal, obedience to malicious data, format error, and parse failure. Keep input hashes and outputs rather than labeling failures merely unstable.

A business page represents structured checks on synthetic input and output. Contextual stock photo, not an IBM interface, real data, or security proof. Image: Unsplash.
A business page represents structured checks on synthetic input and output. Contextual stock photo, not an IBM interface, real data, or security proof. Image: Unsplash.

How to interpret the published benchmark

The article reports author-run static prompt-injection benchmarks. They explain design and experimental results but do not prove the safety of your prompt, tools, parser, and service in production.

Even after synthetic tests pass, continue with independent security review, least privilege, action confirmation, logging, and rollback. This article downloaded no model and reproduced no benchmark.

FAQ

Does LLMON eliminate prompt injection?

That conclusion does not follow from a preview and author benchmarks. Structure is one control among permission and security layers.

Can it be connected directly to production RAG?

Not from this evidence. The preview is for exploration and evaluation, so begin with synthetic data in isolation.

Fact evidence

  • IBM released experimental Granite 4.1 LLMON preview models in 3B and 8B sizes on November 12, 2026. — IBM Granite on Hugging Face — LLMON markup language in Granite 4.1
  • LLMON separates instructions, data, and requirements with explicit structured blocks. — IBM Granite on Hugging Face — LLMON markup language in Granite 4.1
  • The article reports author-run static benchmark results for prompt-injection resistance. — IBM Granite on Hugging Face — LLMON markup language in Granite 4.1
  • The preview models are intended for exploration and evaluation rather than production deployment. — IBM Granite on Hugging Face — LLMON markup language in Granite 4.1

Related tools and reading

These production pages support the same decision, workflow, or safety boundary.

Conclusion

LLMON offers an experimental way to make instruction and data boundaries explicit, but markup alone cannot guarantee safety. Validate with synthetic cases, fixed controls, and four result classes before stricter independent review.

What this means for everyday users

The article reports author-run static prompt-injection benchmarks. They explain design and experimental results but do not prove the safety of your prompt, tools, parser, and service in production.

Related reading

GitHub Updates Secret Scanning Coverage: Recheck AI Workflow Credentials

GitHub published Secret Scanning coverage updates on August 7, 2026. Expanded detectors and coverage can help teams find credentials in AI-generated code, automation, MCP services, and CI, but detection is not a guarantee that every secret will be found. The practical job is to inventory repository, branch, workflow, and local environment sources, confirm organization policy, and respond to an alert by revoking or rotating the credential before investigating history and impact. Bot and app permissions should be minimal, and a small incident drill should verify notification, ownership, deadlines, and postmortem steps. Treat scanning as the discovery layer of a credential lifecycle, not as permission to put secrets in prompts or logs.

AWS Adds Query-Aware Compression to Reduce Bedrock RAG Context Cost

Review the official scope, availability, ordinary-user task, permissions, cost, review, and rollback checks for AWS Adds Query-Aware Compression to Reduce Bedrock RAG Context Cost.

GitHub Code Scanning Adds AI Application Risks: Check Prompt Injection, Disclosure, and Excessive Agency

GitHub has introduced a public preview of code scanning for AI applications and agent vulnerabilities. The detector covers risks such as indirect prompt injection, sensitive information disclosure, and excessive agency, and runs on pull requests alongside existing CodeQL analysis when the required scans are enabled. It flags code paths where untrusted content can reach a model that has access to tools, data, or business systems. This article turns the announcement into a focused verification workflow: confirm prerequisites, create harmless test cases, review alerts in context, reduce permissions, rerun scans, and preserve human approval. The preview requires GitHub Advanced Security and is not currently available on GitHub Enterprise Server.

From Chat Boxes to Personal AI Companions: AI Assistants Are Entering the Desktop Execution Era

AI assistants are moving from answering questions toward continuing real tasks. AI agents, MCP tool ecosystems, personal memory, and local workbenches are pushing this shift together. For users, the real value is not another chat box, but less repeated context setup and more continuity from thinking to doing.

ChatGPT Adds an Accounting Plugin: Keep Source, Variance, and Approval Tables for Month-End Work

OpenAI announced the ChatGPT Accounting plugin on November 18, 2026. It can help reconcile transactions, draft invoice entries, and prepare month-end reports from records and templates, while Enterprise workspaces still require administrator approval before installation. This article does not provide personalized accounting or financial advice. Instead, it supplies three control tables: a source table for files, periods, and versions; a variance table for every unmatched amount; and an approval table that names who may confirm, export, or roll back a change. Bank access, payments, credentials, and final posting remain human-reviewed steps. The plugin output is a reviewable draft, not a closed accounting period or evidence that the underlying records are complete.

GitHub Copilot Evals Enters Preview: Regression-Test Skills, Agents, and MCP Changes

GitHub launched Copilot Evals in public preview on November 17, 2026. It creates repeatable tests for skills, custom agents, and MCP tools and lets users compare runs before and after a customization change. This guide defines a minimum regression set: choose three realistic tasks, then specify the input, allowed tools, success criteria, and prohibited behavior for each. Preserve at least one baseline run and one post-change run, including the result, execution path, failed checks, and human review. Merge a customization only when the change preserves the original task, permission boundary, and evidence chain. A single passing answer is not enough because model output can vary, and an automated check cannot decide every domain-specific quality or safety question.