AI-Discovered Vulnerabilities Need a Faster Disclosure Contract
Turn AI-assisted vulnerability discovery into explicit validation, response, patch, and release states instead of applying one universal deadline.
中文AI engineering, technical decisions and real applications. A blog by KK.
58 articles
Turn AI-assisted vulnerability discovery into explicit validation, response, patch, and release states instead of applying one universal deadline.
中文Audit state, event recovery, permissions, credentials, sandboxes, and data boundaries before moving a long-running agent between managed APIs.
中文Audit Claude Code permission files as sensitive data across command approval, local settings, Git history, build artifacts, and credential rotation.
中文FlashProve can bind checked tokens to a committed integer model. This audit separates model identity, output integrity, safety, and implementation gaps.
中文Anthropic's 123-library experiment shows that LLMs can scale proof writing. A deployment assurance manifest shows what was proved, what ships, and what remains exposed.
中文A caller's knowledge is not identity proof. Use a five-step protocol to exit the call, reopen a trusted channel, contain losses, and preserve evidence.
中文WhatsApp's hidden chatbot instructions show why RAG systems must separate source trust from instruction authority and audit model-visible content.
中文Claude traced NIST curve seeds to lrand48 timestamps. This evidence audit separates reproducible mechanism, historical attribution, and intent.
中文A study of 17,211 citations shows how source preference can amplify ordinary web publishing into an AI-search security path, with important causal limits.
中文Evaluate AI model distillation defenses with three ledgers: capability leakage, legitimate-user friction, and false-positive harm.
中文AI self-replication needs a five-layer evidence contract covering capability, resources, persistence, propagation, and containment.
中文Anthropic OSS Scanner turns raw AI vulnerability reports into a maintainer-controlled queue. Here is the verification contract it needs.
中文Network Time Security authenticates NTP, but an AI agent audit trail also needs synchronization health, explicit time semantics, and evidence binding.
中文Automated red teaming creates lasting value when attacks become verified training data, isolated regression sets, and measurable defense improvements.
中文Gemini 4 Argon shows how identity, purpose, capability, environment, evidence, and revocation can turn model access into a security control.
中文OpenAI's Australia incident shows how AI agent governance must link task authority, runtime evidence, disclosure, remediation, and closure.
中文Conversational AI privacy requires one data-flow audit across web, mobile, backends, trackers, shared links, identifiers, and deletion controls.
中文Assess GLM-5.3 cyber risk through capability, access, attacker uplift, safeguards, and observed misuse as separate evidence layers.
中文OpenAI's DNS egress incident shows why AI agent sandbox containment must cover resolvers, transitive services, telemetry, and automatic shutdown.
中文Private AI Glasses Need a Verifiable Data-Flow and Deletion Contract
中文Audit coding agent data flows before login with canary repositories, auth-state diffs, filesystem traces, egress capture, and deletion evidence.
中文LLM relays can read prompts and rewrite tool calls. Use this evidence-based audit contract before production data or agents cross a proxy.
中文Audit state, authority, recovery, observability, and cost before migrating production agents to OpenAI's managed Codex harness.
中文The DseWiki incident shows why read-only web access must be defined by side effects, not HTTP method names. Here is what the public evidence proves.
中文Persistent AI agents need separate contracts for wake timing and action authority. This checklist makes long-running agents safer.
中文A sponsored label identifies the ad but cannot prove that advertising leaves ChatGPT answers unchanged. Here is a repeatable audit contract.
中文MCP agent identity is becoming a protocol priority. This guide separates stable authorization from draft workload identity, delegation, and DPoP work.
中文A code and policy audit of PhoneBuddySDK shows why mobile AI agents need an in-process runtime, plus the evidence still required to ship.
中文AI safety controls need independent logs, synthetic probes, expiring exceptions, and end-to-end evidence when blocking is disabled.
中文A local-vision, cloud-reasoning architecture keeps home images private while using calibrated escalation, event logs, and reversible actions.
中文WhatsApp Scam Alert shows how on-device ML, confidential analytics, transparency logs, and fail-closed verification can protect E2EE messaging.
中文A 2026 study decoded 315,320 client-side reasoning blocks. The durable lesson is that encryption needs identity, context, expiry, and replay controls.
中文Muse Glimmer fits a 30B agent model into 24 GB. The real deployment contract spans memory, latency, tools, state, permissions, and recovery.
中文OpenAI has formally designated GPT-6 Astra as Critical for cybersecurity. Here is the evidence, safeguard case, and residual risk.
中文A capability matrix for agent security hooks that separates visibility, evidence quality, activation, failure behavior, and enforcement.
中文California DROP reaches 600+ data brokers. Use status, matching, deadlines, and follow-up checks to build a verifiable deletion workflow.
中文AI slop CVEs can trigger harmful patches. Require provenance, source, reachability, reproduction, and fix evidence before auto-remediation.
中文Securely erase an SSD before selling it with risk-based sanitization, completion checks, validation, and a reusable evidence record.
中文Chrome's AI vulnerability pipeline shows why security teams should measure verified, shipped, and installed fixes instead of counting discovered bugs.
中文A Word AI-worm PoC shows how hidden prompts can reach Copilot-generated files. Use context quarantine, provenance, and release gates.
中文The Hugging Face incident shows how persistence compounds small weaknesses. Long-horizon AI safety needs trajectory-level control.
中文Build a cheap behavioral drift sensor for LLM APIs, then use an evidence ladder to keep detection separate from identity and root-cause claims.
中文ChatGPT Health can read connected medical records across conversations.
中文AI agent sandbox persistence varies by lifecycle action. Use this five-layer model to verify files, memory, volumes, connections, and side effects.
中文AI coding agents may execute dependencies before review.
中文MCP Elicitation turns missing input into recoverable control flow. Learn form vs URL mode, state handling, security boundaries, and retries.
中文Browser agent detection now needs three layers: behavioral risk signals, signed identity, and explicit authorization for high-risk actions.
中文Anthropic's CJS scores AI jailbreak severity across four axes. This guide compares CJS, JEF, and CVSS, then separates severity from risk.
中文"OpenAI Advanced Account Security: How Phishing-Resistant Login and Enhanced Protections Work"
中文"Codex Windows Sandbox: Inside OpenAI's Secure Coding Environment Architecture"
中文E2EE messages are only encrypted in transit. Once backed up to the cloud, encryption breaks.
中文Enterprise AI adoption in 2026 hinges on one question that no benchmark can answer: which platform survives your security audit?
中文Meta processes over 100 billion end-to-end encrypted messages daily across WhatsApp and Messenger.
中文"OpenAI published a five-pillar cybersecurity action plan with $10M in API credits and a dedicated GPT-5.4-Cyber model.
中文"OpenAI released its first open-weight PII detection model under Apache 2.0.
中文"Anthropic assembled 12 tech giants and built a cybersecurity AI model too dangerous to release publicly.
中文"The Enterprise AI Security Landscape in 2026: From Guardrails to Trusted Access Programs"
中文On March 27, 2026, Anthropic leaked approximately 3,000 unpublished assets due to a CMS configuration error.