← All posts

Problem-Aware

249 posts on problem-aware.

OpenAI Cannot Rule Out Critical Cyber Capability In Astra

OpenAI cannot rule out Critical cyber capability in Astra, an upcoming model, under its own Preparedness Framework. Four of the five safeguards it names are the lab's internal security programme: isolated testing, weight encryption, sandboxing, and chain-of-thought monitoring. One item survives deployment unchanged, restricted network and tool access, and it becomes a per-request authorization decision on your side of the boundary the moment the model ships. This piece names what produces the audit record when someone asks which instruction produced which outbound call.

ai-securityllm-securityagentic-aipolicy-enforcementai-governanceinline-enforcement
Read post →

AI Red Teaming Research Tracker: Changes to Add to Your Regression Suite

A dated AI red teaming research tracker for security engineers who maintain regression suites. The October 2026 entry maps Microsoft PyRIT 1.1.0 and the NIST adversarial machine learning taxonomy to concrete test-suite changes, expected results, rerun triggers, and the HTTP routes that can produce usable enforcement evidence.

ai-securityllm-securityprompt-injectioninline-enforcementaudit
Read post →

Embedding Leakage: 92% Text Recovery From Vectors

Embedding leakage can expose source text: a 2023 vec2text paper recovered 92% of short inputs exactly from embeddings and recovered patient names from clinical notes. The result changes how teams should classify and access-control vector stores in RAG systems.

ai-securityllm-securitycybersecurityai-governancezero-trust
Read post →

Calendar Invite Prompt Injection: Treat Event Text as Untrusted Input

A calendar invitation can carry instructions that an AI assistant later reads as context. The attack begins before the user asks a question, then fires when the assistant retrieves the event. Security teams need untrusted-source labels, action authorization, destination limits, human approval for consequential steps, and a trace that joins the event to every model decision.

prompt-injectionllm-securityagentic-aiai-securitypolicy-enforcementaudit
Read post →

AI Security Statistics Need Their Denominators Attached

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.

ai-securityshadow-aicybersecurityai-governanceaudit
Read post →

Prompt Injection Benchmarks: AgentDojo, BIPIA, CyberSecEval

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.

prompt-injectionllm-securityai-securityagentic-aiarchitecture
Read post →

Does Otter AI Train on Your Data? Privacy and Enterprise Controls

Otter states that it trains proprietary AI technology on de-identified audio recordings and on transcriptions that may contain personal information. Its Enterprise page advertises an organization-level AI model training opt-out available through an account manager. Buyers should separate that opt-out from transcription processing, retention, sharing, and the terms governing aggregated or de-identified data.

ai-securityai-governancedata-loss-preventionpolicy-enforcementcloud-securityaudit
Read post →

Does Grammarly Train on Your Data? Account Defaults and Controls

Grammarly can use content for model training and product improvement, but the default depends on the account and purchase route. Individual accounts start with training on. Sales-purchased Business, Pro, Education, and Enterprise accounts start with it off. Website-purchased multi-user teams start with it on until an administrator changes the organization setting. Training control also differs from tailored assistance.

ai-securityai-governancedata-loss-preventionpolicy-enforcementcloud-security
Read post →

Does Codex Train on Your Data? Cloud, Workspace, and API Rules

Codex data-use rules depend on the route. Codex Cloud inspects selected repositories and executes tasks in OpenAI-managed workspaces. Business, Enterprise, and Edu workspace data is excluded from model training by default. Direct OpenAI API data is also excluded unless the customer opts in, while retention depends on abuse-monitoring settings and endpoint state. A useful approval names the execution surface, account, repository, retention control, connected services, and training rule.

ai-securityai-governancedata-loss-preventionpolicy-enforcementcloud-securityaudit
Read post →

Does Perplexity Train on Your Data? Consumer, Enterprise, and API Rules

Perplexity applies different model-training rules to consumer accounts, Enterprise products, and its API. Free, Pro, and Max accounts have AI data retention enabled by default but can opt out for future data. Enterprise customer content and API customer content are excluded from model training under current terms. Training, service processing, retention, feedback, and anonymized usage data remain separate review questions.

ai-securityai-governanceperplexitydata-loss-preventionpolicy-enforcement
Read post →