Skip to main content
HikmaAI

Ideas and evidence for the agentic era.

Research, product thinking and practical guidance on securing AI agents in production — from prompt injection and MCP risk to runtime enforcement, governance and sovereign AI.

Read the latest

Read the latest

An illuminated path approaches a secure government data portal.
Research

An AI agent entered Australia's Medicare statistics portal. What should businesses learn?

In June 2026, an experimental OpenAI agent gained unauthorised access to an Australian government statistics service. The case shows why access and incident reporting need clear limits.

Incident: June 2026

Mauro MeddaCo-Founder & CTO · HikmaAI
4 min read
A cyan path passing through layered glass, representing an action and its recorded decisions.
Governance & Evidence

When an AI agent takes action, what should you record?

A transcript shows what an agent said. A useful action record shows who gave it authority, what it tried to do, which rule applied, and what actually happened.

Mauro MeddaCo-Founder & CTO · HikmaAI
4 min read
A bright controlled path runs between two dark research and production rooms.
Governance & Evidence

A model may gain stronger cyber skills. What changes for its security controls?

OpenAI paused some frontier training and tightened research controls after early signs of critical cyber capability. Capability, access, and monitoring must be reviewed together.

Statement: August 18, 2026

Mauro MeddaCo-Founder & CTO · HikmaAI
4 min read
A glass test chamber connected by a lit route to offices beyond it.
Assessment & Red-Teaming

Claude's cyber tests reached real systems. What failed at the boundary?

Anthropic disclosed three cases where cyber evaluations reached real organizations. The test prompt said there was no internet access, but the environment allowed it.

Disclosure: July 30, 2026

Mauro MeddaCo-Founder & CTO · HikmaAI
4 min read
A lit path crosses a glass boundary into a protected data centre.
Research

OpenAI's test agents reached Hugging Face. What failed at the boundary?

A cyber test reached a real production system. The first disclosures point to a package proxy, internet access and stolen credentials as boundaries to check.

Incident: July 2026

Mauro MeddaCo-Founder & CTO · HikmaAI
4 min read
Two illuminated data buildings connected by a single cyan line across a glass bridge.
Agentic Security

A remote MCP connection is an access decision. What should you check?

Google Cloud's new remote MCP server lets external agents reach cloud tools. The useful security question is which identity, permissions, and actions that connection carries.

Announcement: June 30, 2026

Mauro MeddaCo-Founder & CTO · HikmaAI
3 min read
Blue and amber light passes through rows of glass records in a dark archive.
Research

LangGraph's checkpointer flaws: when an agent's saved state becomes an attack path

A June 11 research disclosure showed how exposed history filters and unsafe checkpoint loading could be chained. The risk depends on specific self-hosted configurations.

Research published: June 11, 2026

Mauro MeddaCo-Founder & CTO · HikmaAI
4 min read
A small untrusted module sends an amber signal across glass toward a protected server.
Agentic Security

A malicious coding-agent skill tried to steal a GitHub token. What should teams check?

Datadog found a public skill that requested a GitHub token. Its tests show why a model's refusal may come too late when skill commands run before the model reads them.

Research published: May 11, 2026

Mauro MeddaCo-Founder & CTO · HikmaAI
4 min read
Security and compliance leaders reviewing identity audit materials in a conference room.
Agentic Security

Every AI agent needs an identity and an access review.

An agent can use credentials, call tools, and move data. Security teams need to know what it can access, which actions are controlled, and what evidence remains.

Mauro MeddaCo-Founder & CTO · HikmaAI
3 min read
Engineers investigating dependency inventory during a late-night incident response session.
Research

The LiteLLM breach and the inventory you don't have yet.

Compromised LiteLLM packages were published on March 24, 2026. A current agent and dependency inventory helps teams find where an affected package was used and what needs review.

Mauro MeddaCo-Founder & CTO · HikmaAI
5 min read
Compliance and security professionals preparing audit materials in a conference room.
Governance & Evidence

MCP servers and the audit questions your agents must answer.

An April report raised a concern about MCP server security. The practical question is how to assess connected tools and keep records of the actions an agent takes.

Mauro MeddaCo-Founder & CTO · HikmaAI
4 min read

Editorial categories

Agentic Security

How the attack surface changes when AI can call tools, use credentials and take actions.

Runtime Protection

Practical thinking on gateways, guardrails, egress control, PII redaction and containment.

Assessment & Red-Teaming

How to test agents, models, MCP servers and skills before real attackers do.

Governance & Evidence

Auditability, HikmaScore™, policy, compliance outputs and the board-level questions security teams need to answer.

Sovereign AI

Why deployment location, data boundaries and operational control matter for European and regulated organizations.

Research

Experiments, datasets, papers and technical findings from HikmaAI and research collaborators.

Proven, not promised.

Book a 30-minute live demo