Humanbound website
Blog

Blog

AI security insights, research, and product updates from the Humanbound team.

Techy cover for the Humanbound Agent Attack Scenario Library, showing a test_pack.yaml code card that maps an agent failure to an OWASP Agentic category and the guardrail that closes it, with Agent Goal Hijack, Tool Misuse, and Memory Poisoning tags.
AI SecuritySep 3, 2026

Beyond AI Security: The Rise of AI SecOps: Splunk + Humanbound

Splunk is where most SOCs already live. Here's how Humanbound streams AI agent security findings into Splunk as structured, HMAC-signed webhooks, so agent security gets the same real-time alerting and incident response as everything else the SOC already handles.

SA
Sofia Aliferi
Techy cover for the Humanbound Agent Attack Scenario Library, showing a test_pack.yaml code card that maps an agent failure to an OWASP Agentic category and the guardrail that closes it, with Agent Goal Hijack, Tool Misuse, and Memory Poisoning tags.
AI SecuritySep 3, 2026

Beyond AI Security: The Rise of AI SecOps: Microsoft PyRIT + Humanbound

PyRIT is Microsoft's open-source, research-grade red teaming framework. Here's how a point-in-time PyRIT engagement and Humanbound's continuous, compliance-mapped monitoring plug into each other, including how PyRIT's findings can train the Humanbound Firewall directly.

SA
Sofia Aliferi
Techy cover for the Humanbound Agent Attack Scenario Library, showing a test_pack.yaml code card that maps an agent failure to an OWASP Agentic category and the guardrail that closes it, with Agent Goal Hijack, Tool Misuse, and Memory Poisoning tags.
AI SecuritySep 3, 2026

Beyond AI Security: The Rise of AI SecOps: Promptfoo + Humanbound

Promptfoo and Humanbound aren't competing for the same slot. Here's how Promptfoo's CI red teaming and Humanbound's continuous, compliance-mapped monitoring plug into each other, including how Promptfoo's scan results can train the Humanbound Firewall directly.

SA
Sofia Aliferi
Techy cover for the Humanbound Agent Attack Scenario Library, showing a test_pack.yaml code card that maps an agent failure to an OWASP Agentic category and the guardrail that closes it, with Agent Goal Hijack, Tool Misuse, and Memory Poisoning tags.
AI SecurityAug 31, 2026

Agent Security Debt: Nobody Is Trying to Break Your AI Agent until It Ships

In 2017 I got a CVE for an unencrypted smart bulb. Nine years later, AI agents are shipping with the same gap: nobody tried to break them before launch. Here's how to break your own agent this afternoon, before someone else does.

Ayan Pahwa
Ayan Pahwa
Techy cover for the Humanbound Agent Attack Scenario Library, showing a test_pack.yaml code card that maps an agent failure to an OWASP Agentic category and the guardrail that closes it, with Agent Goal Hijack, Tool Misuse, and Memory Poisoning tags.
AI SecurityAug 20, 2026

The Agent Attack Scenario Library: A Community Reference, Mapped to OWASP

A community library of agent attack scenarios, each mapped to the OWASP Top 10 for Agentic Applications and paired with the guardrail that closes it. Open, credited, and free to contribute to.

SA
Sofia Aliferi
Infographic titled 'The Humanbound Firewall: A Multi-Tier Defense for AI Agents' on a dark chalkboard-style background with white and orange text. The layout is divided into five sections. Top-left, 'Why Single-Layer Guardrails Fail' highlights three problems: The Context Gap, Multi-turn Blindness, and The Novelty Lag, each with hand-drawn icons. Below it, 'The 4-Tier Defensive Stack' shows a layered pyramid diagram with four tiers: Tier 0 (Sanitization), Tier 1 (Generic Attack Detection using ensemble detectors like DeBERTa), Tier 2 (Domain-Specific Classifiers with normal traffic and adversarial pattern models plus a streaming LLM), and Tier 3 (The LLM Judge). Arrows show data flowing upward through the tiers. Top-right, 'The Continuous Test-and-Defend Loop' illustrates a cycle where the humanbound CLI/SDK engine runs adversarial tests producing transcripts, which become labeled training data. Below it, 'The Feedback Loop' shows how that data refreshes the Tier 2 classifier inside humanbound-firewall, which protects the live agent in production. Bottom-right, 'Open Source vs. Managed Platform' contrasts the Apache-2.0 licensed open-source firewall (offering transparency and auditability) with the managed platform for CISOs (providing continuous testing, posture scoring, compliance evidence, and cross-deployment intelligence). A NotebookLM logo appears in the bottom-right corner.
AI SecurityMay 11, 2026

Why we open-sourced humanbound-firewall

We released humanbound-firewall under Apache-2.0. A multi-tier runtime defense for AI agents, with each layer inspectable, escalating on uncertainty, and trainable on your own adversarial test data.

DG
Demetris Gerogiannis
Infographic titled "The Great AI Security Divide: AI4Sec vs. Sec4AI" comparing two distinct AI security categories across primary targets, buyers, threat models, outputs, and key vendors, with a jagged orange lightning crack down the middle separating the two worlds. Includes the Mythos example, three vendor qualification questions, and a "Budget vs. Product" future outlook.
AI SecurityApr 27, 2026

AI Security Means Two Different Things. Mythos Just Made That Visible.

AI security maps to two different markets: AI for security (AI4Sec) and security for AI (Sec4AI). The Claude Mythos Preview made the distinction unmissable. Here is how to tell which one you are actually buying.

KS
Kostas Siabanis
Diagram showing moderation and policy reasoning as two layers of LLM safety
AI SecurityApr 7, 2026

Beyond Moderation: Why LLM Systems Need a Policy Layer

Moderation APIs catch harm and injection attempts but fail to enforce domain-specific policy. A cross-domain evaluation shows why production LLM systems need both moderation and policy reasoning layers.

SB
Spyros Briakos
Abstract representation of AI agents evolving over time while security tests remain static
AI SecurityMar 17, 2026

Your Agent Passed Its Security Test. That Was Three Weeks Ago.

The security industry is applying shift-left to AI agents. But AI agents aren't deterministic. The gap between testing on deploy and staying secure in production is where risk accumulates.

KS
Kostas Siabanis
Abstract illustration of AI agents navigating security barriers and test harnesses
AI SecurityMar 11, 2026

The Enforcement Illusion: Why AI Agent Security Starts with Testing, Not Walls

The AI agent security market is fragmenting into enforcement, identity, and control plane vendors. But the incident data tells a different story: most agents ship without any adversarial testing at all.

KS
Kostas Siabanis