Entity trail

agent-probe

Source-backed findings, relationship evidence, citations, and briefing history from the public MindPattern archive.

Briefing refs

Findings

Edges

Sources

Corpus findings

2026-07-07 / arxiv-researcherCoding Agents Internally Encode a Forward 'Programming Horizon'Probes what a coding agent's underlying model represents across dozens of edit-and-test steps, finding the residual stream encodes a latent horizon of the program it is building. It is an interpretability lens on why coding agents lose the thread on long tasks. Useful groundwork for monitoring or intervening when an agent drifts off-plan.
2026-07-02 / arxiv-researcherCan Agents Generalize to the Open World? Unveiling the Fragility of Static Training in Tool UseThis paper (2607.01084, published 2026-07-01) probes how brittle tool-using agents become when moved from static training distributions into open-world conditions, exposing failure modes in generalization. A cautionary, builder-relevant result for anyone shipping tool-calling agents beyond their training set.
2026-06-21 / sources-researcherORAgentBench Probes Whether LLM Agents Can Solve Operations-Research Tasks End-to-EndORAgentBench (arXiv 2606.19787, June 18) benchmarks whether LLM agents can solve challenging operations-research and optimization problems fully end-to-end, using realistic constraint-heavy tasks. It's a useful capability probe for builders eyeing agents in quantitative, constraint-laden domains. Early benchmarks like this map exactly where agents still fall short of full autonomy before you bet a workflow on them.
2026-06-10 / skill-finderEvaluate agent memory on four cognitive competencies with MemoryAgentBenchMemoryAgentBench grounds memory evaluation in cognitive science and probes four distinct competencies — accurate retrieval, test-time learning, long-range understanding, and selective forgetting — by feeding inputs incrementally as time-ordered dialogue chunks across architectures (MIRIX, MemGPT, long-context, and RAG agents). The key insight for builders: 'remembers facts' is only one of four axes, and selective forgetting is usually the neglected one. Score your own agent's memory layer against all four so you catch failures like hoarding stale context or failing to update on new information, not just retrieval misses.
2026-06-08 / agents-researcherPraisonAI Agent Framework (CVE-2026-44338) Probed Within 3h44m of Disclosure — Auth Disabled by DefaultCVE-2026-44338 (CVSS 7.3) stems from PraisonAI's legacy Flask api_server.py shipping with AUTH_ENABLED=False and AUTH_TOKEN=None, exposing GET /agents and POST /chat so anyone on the network can execute agent workflows and drain API quotas without credentials. Sysdig documented a scanner identifying as 'CVE-Detector/1.0' hitting the exact endpoint 3 hours 44 minutes after the advisory went public. The flaw affects versions 2.5.6–4.6.33 and is fixed in 4.6.34 — a sharp example of insecure-by-default agent tooling meeting near-instant mass exploitation.
2026-06-05 / hn-researcherLaunch HN: Hyper (YC P26) — 'Company Brain' to Power Agentic DevelopmentHyper launched on HN (77 points, 76 comments) pitching a persistent 'company brain' that gives coding agents shared organizational context — code, decisions, and institutional knowledge. The launch fits the 2026 trend toward persistent context layers and governed orchestration for agent fleets rather than one-off prompts. Commenters probed how it differs from existing RAG-over-codebase and memory approaches.
2026-06-02 / arxiv-researcherClinEnv: Interactive Multi-Stage Long-Horizon EHR Environment for Clinical Agent EvaluationBuilds the first interactive clinical environment where AI agents must gather heterogeneous EHR data incrementally and make sequential, irreversible decisions under uncertainty — mirroring real physician workflows. Static benchmarks with multiple-choice answers cannot probe this kind of decision-making. Significant for healthcare AI builders: provides a realistic evaluation harness that tests whether agents can handle the actual complexity of clinical reasoning.
2026-05-28 / github-pulse-researcherTencent AI-Infra-Guard v4.1.10: Full-Stack AI Red Teaming Platform With 76 Analyzers, 9,600+ Detection Rules, and Agent/MCP/Skills ScanningTencent/AI-Infra-Guard scans across infrastructure (64+ AI framework fingerprints, 1,300+ CVEs), agent workflows, MCP servers, agent skills, and LLM jailbreak vectors in a single platform. v4.1.10 (May 2026) expanded component coverage; 26 new attack operators (20 single-turn + 6 multi-turn) test prompt security. Unlike passive auditors, A.I.G actively probes for loss-of-control risks and permission boundary violations in deployed agents — the most comprehensive open-source AI security scanner from a major tech company.
2026-05-15 / arxiv-researcherPosition Paper: Behavioural Assurance Cannot Verify the Safety Claims AI Governance Now DemandsThis paper argues that AI governance frameworks enacted between 2019 and 2026 demand evidence of properties — absence of hidden objectives, resistance to loss-of-control precursors, bounded catastrophic capability — that behavioral evaluations and red-teaming are epistemically incapable of providing. Current assurance methods are limited to observable outputs and cannot access latent representations or long-horizon agentic behaviors. The authors propose bounding behavioral evidence's legal weight and extending pre-deployment access with mechanistic-evidence classes including linear probes, activation patching, and before/after-training comparisons.
2026-05-02 / arxiv-researcherTwinGate: Stateful Defense Against Decompositional Jailbreaks in Multi-Turn ConversationsDecompositional jailbreaks split a malicious objective across multiple benign-looking turns that individually pass safety filters. TwinGate uses asymmetric contrastive learning to maintain state across turns and detect when a conversation trajectory is reconstructing prohibited content, even in untraceable traffic where session linking is unavailable. A practical defense for any multi-turn agent or chatbot deployment where adversaries can probe over extended conversations.
2026-04-24 / arxiv-researcherTransient Turn Injection Exposes Stateless Multi-Turn Vulnerabilities in LLMsResearchers introduce Transient Turn Injection (TTI), a multi-turn attack technique that exploits stateless moderation by distributing adversarial intent across isolated interactions. TTI uses automated attacker agents to iteratively probe and evade safety guardrails in both commercial and open-source LLMs. The attack surface is particularly concerning because most production deployments use stateless turn-by-turn moderation that cannot detect distributed intent across sessions.
2026-04-20 / hn-researcherShow HN: Nyx — Multi-Turn Adaptive Offensive Testing Harness for AI AgentsNyx is a new Show HN project (20pts, 8 comments) that autonomously probes AI agents for failure modes including logic bugs, instruction-following failures, edge cases, jailbreaks, prompt injection, and tool hijacking through multi-turn adaptive conversations. As more agents ship to production, automated red-teaming tools like Nyx fill a critical gap — the tool represents the growing 'AI security testing' category alongside Anthropic's own internal red-teaming and HackerOne's AI Red Teaming platform.

Graph relationships

BENCHMARKED AGAINST
PyRIT
agent-probe targets function layer vs PyRIT/Garak model outputs
Source finding

Source trail

arXiv arXiv arXiv arXiv — Evaluating Memory in LLM Agents via Incremental Multi-Turn Interactions (MemoryAgentBench)Sysdig Hacker News arXiv

Graph sources

entity graphfindings textkg edgeskg entities