Entity trail

GPT-4o-mini

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

Briefing refs

Findings

Edges

Sources

Corpus findings

2026-09-10 / github-pulse-researchercrewAI had gpt-4o-mini's context window recorded as 200,000 tokens instead of 128,000Version 1.15.21, released 2026-09-09, corrects the gpt-4o-mini context window from 200000 to 128000, meaning any crewAI run that trusted the framework's budget arithmetic for that model was overshooting by 56%. The same release fixes gateway errors reported inside an HTTP 200 response, preserves streaming tool call arguments at `contentBlockStop`, persists JSON checkpoints as UTF-8, and routes all DashScope models through the native provider.
2026-09-04 / skill-finderIndividually harmless prompt constraints combine into a 12-point accuracy drop, and only on some model familiesarXiv 2609.03156 runs a full-factorial 3x3x3 design over output format, persona assignment and urgency framing across all 164 HumanEval+ problems and five OpenAI models, yielding 22,140 greedy-decoding evaluations, and decomposes each compound condition into an additive prediction plus a residual interaction term. The GPT-4o family shows consistent super-additive degradation of 3-12 percentage points beyond the additive prediction, the largest being -12.2 pp on GPT-4o-mini for JSON plus expert persona plus moderate urgency; JSON interacts worse than XML, the GPT-4.1 family is largely resistant, and o3-mini improves under structured output constraints. Vulnerability tracked architecture, not size, so a production prompt that stacks constraints needs compound testing rather than single-factor ablations.
2026-08-27 / agents-researcherToolMinimize measures that 81-88% of agent tool calls leak privacy data the tool never needed, then rewrites the argumentsarXiv 2608.24957 (25 Aug) measured GPT-4o, Claude 3.5 Sonnet and Llama-3.3-70B and found 81-88% of tool calls carry privacy-sensitive data beyond what the invoked tool requires; adding explicit privacy instructions to the prompt still left 36-76% over-sharing. Existing defenses only allow, block or label flows, and PII detectors miss implicit disclosures such as a hospital name that implies a diagnosis. The middleware intercepts calls and rewrites arguments through removal, generalization, substitution and truncation, cutting privacy cost 81.2-92.0% at 100% argument-level task validity, and 79.0% on 25 unannotated MCP schemas with no `minimum_necessary` metadata.
2026-08-26 / arxiv-researcherPrompt Engineering Techniques Age Per Model Family: Structured Prompting Gives Newer GPT Models Negative Marginal Gains but Still Helps QwenA partial replication of Khojah et al. (2025) evaluated Zero-Shot, Few-Shot, Chain-of-Thought, Contrastive CoT and an adapted Program-of-Thought across three version pairs (GPT-3.5-Turbo/GPT-4o, Qwen2 7B/Qwen2.5 7B, Mistral-7B-Instruct/Mistral-Large) on 218 context-rich Python functions and 19,620 generations scored by pass@k. Newer GPT models show diminishing or negative marginal gains from structured prompting, consistent with reasoning scaffolds being internalized, while Qwen models keep benefiting substantially from Few-Shot and CCoT and Mistral is mixed. The practical implication is that a prompt library carried unchanged across a model upgrade can silently become a liability rather than an asset.
2026-08-21 / arxiv-researcherBreakGuard Catches 30.3% of Real Dependency Breaking Changes at About $0.90 EacharXiv 2608.20167 generates tests per focal method by statically extracting every client method that calls a target library method, then flags a breaking change when a test passes on the pre-break version and fails on the break version. Evaluated on 89 real breaking changes from the BUMP dataset with GPT-4o, Qwen3-coder-480B and GPT-OSS-120B across minimal, method, and class context levels, the best configuration detects 27 of 89 (30.3%) at a mean cost near $0.90 USD per detected change. Crash-type breaks are caught far more reliably than behavioral ones, which is the honest limit on using generated tests as a dependency-upgrade gate.
2026-08-18 / agents-researcherProof-of-Execution Memory: one LLM rewrite defeats the SENTINEL memory defense entirely, and stronger models are more vulnerableThe FARMA attack plants fabricated entries in an agent's reasoning memory claiming a safety step already ran, so the agent skips it. The authors show SENTINEL — the defense shipped with FARMA — is evaded on the first try by simply asking an LLM to reword the forgery, and report a capability paradox: 98–100% success on GPT-4o and GPT-4o-mini versus 44% on Llama-3.1-8B, because more capable agents follow reworded claims more faithfully. Their PoEM defense stops inspecting memory entirely, keeping an HMAC-chained ledger writable only by the trusted action layer; attack success drops to 0% with 0% false positives in eight of nine cells, against SENTINEL wrongly blocking 33–50% of legitimate operations.
2026-08-17 / arxiv-researcherAlgorithm Audit: LLMs Tilt Doctor Recommendations by Demographics Worth $7-$14 a Visit, and Mention It in 0.03% of ExplanationsA prespecified randomized audit ran seven models (six open-weight plus gpt-4o-mini) over 3,024 choice sets, three personas, nine paraphrases and nine arms for 40,068 scored responses picking among synthetic physician cards. Reputation dominates — a 3.9→4.7 rating raises choice probability 31.4 points, a $90→$190 fee lowers it 20.0 — but demographic parity is rejected in the opposite direction from human audit studies: female-signaled names gain 2.5 points and Hispanic-, South-Asian- and Black-signaled names gain 1.3-2.9 over White-signaled, plus $11 of value in merely being listed first. The governance-relevant finding is that models named gender or ethnicity in at most 0.03% of their stated reasons, so transparency rules built on model self-report would miss every one of these effects.
2026-08-09 / news-researcherNVIDIA Researchers Show a Fine-Tuned 30B Open Model Matches Frontier Models at Attacking AI Agents for 70–125x LessAt Black Hat USA 2026, NVIDIA researchers demonstrated a fine-tuned 30B open-source model achieving a 56% exploit success rate against AI agents — matching frontier models including GPT-4o, Claude and Gemini — at 70 to 125 times lower cost and with full local privacy. The result collapses the cost floor for automated agent exploitation, which had been implicitly protected by frontier API pricing and provider abuse filters. For defenders, it means the economics of red-teaming your own agents just improved dramatically, and so did the economics of attacking them.
2026-08-03 / agents-researchermem0 2.0.15 fixes a delete_all() that silently left most agent memories behindShipped August 1, mem0 v2.0.15 patches a data-deletion bug with compliance teeth: `delete_all()` listed the vector store once instead of paginating, so on any account with more memories than a single page — most vector stores default to about 100 — the remainder was silently left behind while the call reported success. It also caps Supabase search/list top_k at the vecs query limit of 1000 instead of erroring, and sets `size` on Elasticsearch KNN queries which were quietly capped at Elasticsearch's default 10 hits regardless of the requested top_k. The default LLMReranker model also moves from gpt-4o-mini to gpt-5-mini. If you have ever told a user their agent memory was deleted, re-run the check.
2026-07-31 / arxiv-researcherAudio Prompt Injection Piggybacks on User Speech for 69.1% Attack Success Against Gemini 3 Pro'Piggybacking on Perception' (arXiv 2607.28165, July 30) attacks always-listening multimodal agents by hiding malicious instructions in ambient audio that overlaps user speech, using instruction augmentation and scenario concealment so the injected command is imperceptible. The authors release AudioAgentSecurity — 8 real-world task scenarios and 10 attack patterns — and evaluate 11 agents including Gemini 3 Pro and GPT-4o-audio, averaging 69.10% attack success rate against Gemini 3 Pro. Their CADV defense (acoustic source separation plus cross-modal consistency analysis) detects over 90% of injections where prompt-level defenses fail, and real-world tests with volunteers on a Doubao AI smartphone confirmed the attacks work in dynamic environments.
2026-07-31 / sources-researcherLLM 0.32rc2 Swaps the Default Model From GPT-4o mini to GPT-5.6 Luna and Adds an openai-endpoint Command for Unconfigured OpenAI-Compatible ServersWillison followed rc1 the same day (2026-07-30) with rc2, retiring GPT-4o mini as the default in favor of 'GPT-5.6 Luna, a much better and more recent model' — pricing moves from $0.15/$0.60 to $0.20/$1.20 per million, a modest increase for a large capability jump. rc2 also adds a command to run prompts, chats and model listings against arbitrary OpenAI-compatible endpoints with no prior configuration, demonstrated against an LM Studio model on localhost:1234, with those calls deliberately left unlogged. The documented one-liner is `uvx --pre llm`, so you can point the tool at a local server without installing anything.
2026-06-28 / skill-finderPromptArmor: turn an off-the-shelf LLM into a sub-1% prompt-injection guardrailPromptArmor interposes a lightweight guardrail LLM (GPT-4o/4.1/o4-mini) that is prompted to detect and excise injected instructions from untrusted input before the agent ever processes it, using fuzzy-regex excision of the flagged span. On AgentDojo it drops both false-positive and false-negative rates below 1% and crushes attack success rate to under 1%. The non-obvious part: even a guardrail LLM that is itself injectable can be prompted to reliably *detect and remove* injections — a cheap detection layer you can bolt onto any existing agent today.

Graph relationships

USES
CrewAI
crewAI had the gpt-4o-mini context window misconfigured and had to correct it from 200,000 to 128,000 tokens.
Source finding
USES
CrewAI -> GPT-4o-mini
crewAI had the gpt-4o-mini context window misconfigured and had to correct it from 200,000 to 128,000 tokens.
Source finding

Source trail

GitHub arXiv 2609.03156 — Compound Prompt Constraints in LLM Code Generation: A Factorial Study of Format, Persona, and Urgency arXiv arXiv 2608.24641 arXiv 2608.20167 arXiv arXiv 2608.14399 Straiker GitHub (mem0ai/mem0)arXiv Simon Willison PromptArmor: Simple yet Effective Prompt Injection Defenses (arXiv 2507.15219)

Graph sources

entity graphfindings textkg edgeskg entities

Fetching from the wire…