Analysis on enterprise AI governance, inline policy enforcement, agentic AI security, and regulatory compliance.
AI security statistics become useful when the source year, study population, denominator, and control boundary stay attached. IBM measured breach outcomes across 600 organizations, while Verizon measured workforce access to generative AI on corporate devices. This guide separates those populations and turns each figure into a testable security question.
Read post →Microsoft states that Copilot Studio is covered under its HIPAA Business Associate Agreement and permits agents to handle protected health information. That gives healthcare buyers a documented BAA path. The deployment still needs a route inventory, safeguards, access controls, connector review, audit evidence, and confirmation that every service touching PHI remains inside the agreement.
Read post →The Council of Europe Framework Convention on Artificial Intelligence and Human Rights, Democracy and the Rule of Law is a treaty between states, not a direct compliance rulebook for enterprises. Its operative themes include transparency, accountability, remedies, procedural safeguards, risk and impact management, and independent oversight. This guide separates the treaty text from the domestic measures that parties still have to adopt and identifies the records an enterprise should preserve while those measures take shape.
Read post →Compare seven Cyera alternatives by the job they perform: data discovery, non-human identity governance, LLM observability, DLP, or inline AI policy enforcement. Includes a practical fit test for each category.
Article 99 of the EU AI Act sets three penalty tiers. €35 million or 7% of global turnover for prohibited practices. €15 million or 3% for the obligations listed in Article 99(4). €7.5 million or 1% for supplying incorrect or misleading information. The higher of the absolute figure and the percentage applies to undertakings, and the lower of the two applies to SMEs and start-ups under Article 99(6). This piece walks the exact article list behind each tier, the crossover turnover where the percentage starts to bind, and which obligations are actually live after the July 2026 omnibus deferral.
Read post →Seven prompt injection benchmarks are in active use and their headline attack success rates run from 24% to 84.30%: AgentDojo (629 security cases, ETH Zurich, NeurIPS 2024), BIPIA (Microsoft, KDD 2025), InjecAgent (1,054 cases, ACL Findings 2024), Open-Prompt-Injection (USENIX Security 24), CyberSecEval 2 (Meta, 26-41% on every model), Agent Security Bench (ICLR 2025, 84.30% peak), and MCPTox (45 real MCP servers). Every one of those is a static score. When researchers from OpenAI, Anthropic, and Google DeepMind ran adaptive attacks against twelve published defenses, all twelve fell, most above 90%. This piece gives the comparison table, what each benchmark measures, and the design conclusion the numbers force.
Read post →Document-level security in RAG belongs in the retrieval query, before chunks enter an LLM prompt. Index each document with tenant and access metadata, derive the caller identity from a trusted source, apply the filter server-side, preserve source permissions during updates, and record the retrieved document IDs for each model request.
LLM caching can reduce latency and repeated token spend, but semantic and prefix caches create another store for prompts, responses, embeddings, or provider-side context. Security design needs cache scope, tenant isolation, key construction, encryption, expiry, invalidation, and a rule for which data classes may be cached.
Read post →Secrets detection in prompts should identify provider keys, cloud credentials, private keys, database strings, and organization-specific tokens before an HTTP LLM request leaves the application boundary. Pattern matching needs entropy checks, custom signatures, decoding tests, allowlists, and a response plan for any credential already exposed.
Read post →Open Policy Agent can evaluate structured facts about an AI request, including caller role, model route, data classification, region, and requested tool. OPA returns a decision; the gateway still has to enforce it, fail closed, protect policy inputs, version bundles, and write evidence for each request.
Redaction removes or masks sensitive values before an LLM call. Tokenization replaces them with references that an authorized service may later resolve. The reversible path preserves workflow utility and creates a token vault, access-control, and re-identification obligation. Choose per data field and output requirement.
Read post →AI logs sent to a SIEM should record caller identity, model route, policy version, data classification, decision, latency, and correlation IDs. Raw prompts and responses require a separate content decision. OpenTelemetry can standardize model events, while an independent gateway record proves what policy allowed before the call.
Read post →