https://www.oktsec.com/research Incident analysis, reproducible labs and product research on AI agent authorization, MCP security, identity, evidence and software supply chain risk. en Thu, 03 Sep 2026 21:01:49 GMT https://www.oktsec.com/research/agent-governance-gap-in-numbers-2026 https://www.oktsec.com/research/agent-governance-gap-in-numbers-2026 Thu, 03 Sep 2026 12:00:00 GMT Research 65% saw an agent act out of scope, 34% authorize per action, 46% cannot produce a 30 day audit trail. What four August 2026 reports agree on. https://www.oktsec.com/research/agent-incidents-august-2026 https://www.oktsec.com/research/agent-incidents-august-2026 Thu, 03 Sep 2026 12:00:00 GMT Incident analysis Six agent security events between August 26 and September 2, 2026 point at the same failure: authorization at the action, not detection of bad text. https://www.oktsec.com/research/ai-generated-tooling-industrial-control-systems https://www.oktsec.com/research/ai-generated-tooling-industrial-control-systems Fri, 21 Aug 2026 12:00:00 GMT Threat analysis Joint advisory AA26-231A: AI assistance, public scanning data and legitimate S7 tooling compress the path from reconnaissance to operational capability. https://www.oktsec.com/research/when-agents-leave-the-sandbox https://www.oktsec.com/research/when-agents-leave-the-sandbox Thu, 06 Aug 2026 12:00:00 GMT Incident analysis Agents with tools, credentials and internet access took unsanctioned action during an AISI cyber evaluation. The durable control is the action boundary. https://www.oktsec.com/research/mcp-2026-07-28-security-boundary https://www.oktsec.com/research/mcp-2026-07-28-security-boundary Tue, 28 Jul 2026 12:00:00 GMT Protocol analysis The 2026-07-28 MCP specification removes protocol sessions and leaves state integrity, authorization and resource controls to each implementation. https://www.oktsec.com/research/agent-autonomy-security-boundary https://www.oktsec.com/research/agent-autonomy-security-boundary Tue, 21 Jul 2026 12:00:00 GMT Research Production agents run with approval, supervision or guardrails. Each step away from the human changes what identity, policy, evidence and recovery need. https://www.oktsec.com/research/jadepuffer-agentic-ransomware https://www.oktsec.com/research/jadepuffer-agentic-ransomware Tue, 21 Jul 2026 12:00:00 GMT Threat research Sysdig documented what it assesses as the first agentic ransomware operation: a familiar exploit chain, an adaptive loop and a compressed response window. https://www.oktsec.com/research/openai-hugging-face-evaluation-incident https://www.oktsec.com/research/openai-hugging-face-evaluation-incident Tue, 21 Jul 2026 12:00:00 GMT Incident analysis In an internal cyber evaluation, OpenAI models found a path from a constrained research network into Hugging Face production. The containment boundary failed. https://www.oktsec.com/research/how-to-audit-a-claude-skill https://www.oktsec.com/research/how-to-audit-a-claude-skill Fri, 03 Jul 2026 12:00:00 GMT Research Read SKILL.md, bundled scripts, referenced URLs and requested permissions the way an attacker would, before the skill enters your environment. https://www.oktsec.com/research/how-to-secure-an-mcp-server https://www.oktsec.com/research/how-to-secure-an-mcp-server Fri, 03 Jul 2026 12:00:00 GMT Guide What you are actually securing, why authentication is not enough, how to stop prompt injection reaching a tool, and what the latest MCP specification changes. https://www.oktsec.com/research/authorization-telemetry-for-ai-agents https://www.oktsec.com/research/authorization-telemetry-for-ai-agents Fri, 03 Jul 2026 12:00:00 GMT Research How to instrument the authorization decision on every agent tool call and export identity, tool, arguments and policy verdict through OpenTelemetry. https://www.oktsec.com/research/how-exposed-are-mcp-servers https://www.oktsec.com/research/how-exposed-are-mcp-servers Fri, 03 Jul 2026 12:00:00 GMT Research Across six independent studies, the dominant failure is missing authentication and overbroad credentials, not exotic exploits. https://www.oktsec.com/research/what-the-research-shows-defending-agents https://www.oktsec.com/research/what-the-research-shows-defending-agents Fri, 03 Jul 2026 12:00:00 GMT Research A synthesis of the 2025 to 2026 literature: what detection and filtering defenses score, why adaptive attacks break them, and why authorization before the action holds up. https://www.oktsec.com/research/how-agent-security-benchmarks-score https://www.oktsec.com/research/how-agent-security-benchmarks-score Fri, 03 Jul 2026 12:00:00 GMT Research AgentDojo, InjecAgent, ASB, MCPSecBench, MSB and MCPTox measure different attack classes, so their scores are not directly comparable. https://www.oktsec.com/research/anatomy-of-the-agent-tooling-cves https://www.oktsec.com/research/anatomy-of-the-agent-tooling-cves Fri, 03 Jul 2026 12:00:00 GMT Research The critical 2025 to 2026 CVEs in agent tooling share one shape: untrusted input reaches a privileged action through a configuration the tool trusts automatically. https://www.oktsec.com/research/detection-versus-authorization https://www.oktsec.com/research/detection-versus-authorization Thu, 02 Jul 2026 12:00:00 GMT Research Detection watches behavior after it looks wrong. Authorization decides what an agent may do before it acts and verifies what it did after. https://www.oktsec.com/research/agent-identity-stack https://www.oktsec.com/research/agent-identity-stack Wed, 01 Jul 2026 12:00:00 GMT Research AI agent identity in mid-2026, layer by layer: signed Agent Cards, enterprise managed authorization, task scoped research and the missing per instance layer. https://www.oktsec.com/research/evidence-for-agent-work https://www.oktsec.com/research/evidence-for-agent-work Wed, 01 Jul 2026 12:00:00 GMT Research Evidence is the verifiable record of what an agent actually did: which identity ran, which tools it called, with what arguments, under which policy. https://www.oktsec.com/research/the-agent-supply-chain https://www.oktsec.com/research/the-agent-supply-chain Wed, 01 Jul 2026 12:00:00 GMT Research The agent supply chain is every artifact an agent pulls into your environment: packages, MCP servers, skills and the trust behind them. https://www.oktsec.com/research/skills-over-mcp-trust https://www.oktsec.com/research/skills-over-mcp-trust Wed, 01 Jul 2026 12:00:00 GMT Research SEP-2640 proposes discovering skills through MCP resources. Provenance, pinning, review and instruction precedence remain open trust questions. https://www.oktsec.com/research/mcp-identity-layer https://www.oktsec.com/research/mcp-identity-layer Wed, 01 Jul 2026 12:00:00 GMT Protocol analysis MCP moves authorization context from the connection to each request. Security still depends on how every server implements identity and policy. https://www.oktsec.com/research/agent-identity-revisited https://www.oktsec.com/research/agent-identity-revisited Wed, 01 Jul 2026 12:00:00 GMT Research A2A v1.0 signed Agent Cards, MCP enterprise managed authorization and task scoped token prototypes: what is available, proposed and still missing. https://www.oktsec.com/research/what-an-mcp-server-actually-exposes https://www.oktsec.com/research/what-an-mcp-server-actually-exposes Thu, 28 May 2026 12:00:00 GMT Guide A practical, audit grounded map of tool surfaces, credentials, network reach and the trust boundaries that decide your real attack surface. https://www.oktsec.com/research/prompt-injection-is-an-authorization-problem https://www.oktsec.com/research/prompt-injection-is-an-authorization-problem Tue, 12 May 2026 12:00:00 GMT Note Content filters lose the prompt injection arms race. The durable fix is authorization: decide what an agent may do before it acts and verify what it did after. https://www.oktsec.com/research/mcp-supply-chain https://www.oktsec.com/research/mcp-supply-chain Fri, 01 May 2026 12:00:00 GMT Research Every server an agent calls is an implicit trust decision. The operating problem is how to make that decision explicit and find the blind spots that remain. https://www.oktsec.com/research/runtime-policy-enforcement https://www.oktsec.com/research/runtime-policy-enforcement Wed, 01 Apr 2026 12:00:00 GMT Engineering Static scanning stops at deploy. Agent security needs runtime policy that constrains tools, credentials, destinations and privileged actions before execution. https://www.oktsec.com/research/agentpay-hackathon https://www.oktsec.com/research/agentpay-hackathon Wed, 01 Apr 2026 12:00:00 GMT Build note A prototype that puts deterministic security checks in front of payments an AI agent tries to execute, and the product lesson that now shapes Oktsec. https://www.oktsec.com/research/security-product-decision https://www.oktsec.com/research/security-product-decision Sun, 01 Mar 2026 12:00:00 GMT Product essay The most consequential call a technical founder makes is what not to ship. Agent security cannot wait for an incident to become part of the product. https://www.oktsec.com/research/agent-skills-ecosystem-trust https://www.oktsec.com/research/agent-skills-ecosystem-trust Sun, 01 Feb 2026 12:00:00 GMT Field note A practical review model for public agent tooling: permissions, provenance, install paths, version pinning and change over time. https://www.oktsec.com/research/agent-identity-crisis https://www.oktsec.com/research/agent-identity-crisis Sun, 01 Feb 2026 12:00:00 GMT Research The AI agent identity gap, the delegation problem and the five layers it takes to close it before agents can safely act across real systems.