pi v0.99.0
The open-source Pi coding agent builds in MCP server support, a code mode that runs model-written JavaScript in a QuickJS sandbox, and virtual models that choose an underlying model for each request.
Earendil Works
Heft Radar
What Heft is following
The conservation science, funding and AI research we track while we build Heft, newest first. Every entry links to the original. Follow it by RSS.
119 entries
September 2026
Indo-Burma: Second Call for Letters of Inquiry
Small grants of up to US$50,000 and large grants of US$50,000 to US$250,000 for local civil society groups in Cambodia, China, Lao PDR, Myanmar, Thailand and Vietnam. Closes 2 November 2026.
Critical Ecosystem Partnership Fund
Claude Code v2.1.284
This Claude Code release makes Claude Sonnet 5.5 the default Sonnet model, with a 1M-token context window priced at $2 input and $10 output per million tokens.
Anthropic
Assessing safeguards against the illegal wildlife trade on AI platforms
Tests of seven AI platforms across 20 taxa found they blocked illegal wildlife purchases better than a Google search, but safeguards were weakest for corals, tarantulas and orchids.
Conservation Biology
WideSWE: Can Coding Agents Coordinate Changes Across Repositories?
A benchmark of 120 real tasks that each need coordinated changes in more than one code repository. Full success ranged from 10.83% to 42.50% across seven coding-agent configurations.
arXiv
QwenGyre: An Elastic Reinforcement Learning Framework for Training xLong-Horizon Agents
A reinforcement learning framework for agents whose single runs last hours and reach about 700,000 tokens. The authors report NL2RepoBench rising from 52.5% to 58.5% in 48 training steps.
arXiv
Relic: From Multi-Agent Collaboration to Persistent Organizational Capability
Agent teams turn recurring collaboration failures into protocols enforced at runtime. New members followed inherited rules 41.2% of the time as executable bindings, against 34.6% as plain text.
arXiv
Hidden Mortality: Bottom Trawling Challenges the Legal Protection of Vulnerable Marine Species
Standardised trawl surveys off Israel caught about 53 legally protected animals per trawling hour, peaking at 147.7 per hour at 80 m, despite strict spatial and seasonal limits on trawling.
Conservation Letters
ExpVoyager: Direct Experience Navigation for Dynamic Agent Skill Synthesis
Proposes building agent skills on demand: a curator searches an agent's raw past trajectories for what the current task needs and extracts reusable procedures as it goes.
arXiv
Saint Vincent and the Grenadines: Call for Letters of Inquiry for GEF Small Grants Program
Grants of up to US$50,000 for NGOs, community groups and other civil society organisations in Saint Vincent and the Grenadines working on threatened ecosystems and species. Closes 1 November 2026.
Critical Ecosystem Partnership Fund
Codex 0.157.0
OpenAI's Codex coding agent adds the GPT-6 Sol and Luna models, including through Amazon Bedrock, and starts its local background server automatically for eligible interactive sessions.
OpenAI
Legacy Landscapes Fund gets a major boost for landscapes in Africa
Germany, the Rob Walton Foundation and the Wyss Foundation each commit US$100 million, US$300 million in total, to expand the fund's long-term financing of conservation landscapes in Africa.
Legacy Landscapes Fund
Interim GEF CEO statement on the UK Government's pledge to the GEF-9 replenishment
The UK pledges £330 million to the ninth replenishment of the Global Environment Facility, which the GEF says takes total GEF-9 pledges above US$4.3 billion.
Global Environment Facility
30x30 Global Assessment 2026
About 20% of land and 10% of the ocean are designated as protected or conserved, and only about 6.5% of protected areas have reported a management-effectiveness assessment since 2017.
Campaign for Nature
vLLM v0.30.0
The inference engine adds a per-GPU daemon that keeps quantised, sharded weights in memory, so a restarted engine maps them over CUDA IPC without reloading from disk. 762 commits from 315 contributors.
vLLM
Claude Code v2.1.280
Claude Opus 5.5 becomes the default Opus model, with a 1M-token context at $4 input and $20 output per million tokens. The release also stops symlinked paths from slipping writes past approval.
Anthropic
RRSI: Regularized Recursive Self-Improvement of Agent Harnesses
Self-improving agent harnesses can overfit their training tasks. RRSI constrains how edits are proposed and kept, gaining up to 14.1 points in-distribution and 4.7 on five unseen benchmarks.
arXiv
Jev-Mem: System-One-Controlled Agentic Memory for Efficient AI Agents
A fast decision model routes, scores and stops memory retrieval, and a language model writes only the answer. On LoCoMo the design built memory 6.6 times faster and cut query latency by 36.7%.
arXiv
Harness-Zero: Harness Distillation via Agent-as-Harness
Harness-Zero trains a model to do without its specialised agent harness. With the harness removed, task success rose from 23.3% to 44.3%, above the 41.7% reached with the harness attached.
arXiv
DeepSeek Elastic Compute (DSec): A Sandbox Infrastructure for Effective Agentic Training at Scale
DeepSeek describes the sandbox platform behind its agent training, with function-call, container, microVM and full-VM backends. One unit of about 160 nodes serves roughly 3 million sandboxes a day.
arXiv
ZCode uploads your git history; Z.ai holds the only key
A reverse-engineering report finds the ZCode coding agent packs whole workspaces, git history included, into encrypted archives sent to cloud storage, and that two settings toggles did not stop it.
Tokenstead
Jev 1.13
TypeSafe's first System One model returns structured decisions in place of generated text. It is hosted only, priced at $0.042 per million input tokens with output free.
TypeSafe AI
Claude Code v2.1.277
Claude Code now reads AGENTS.md when a project has no CLAUDE.md file, with the choice set under project instructions in its configuration. Bedrock, Vertex and Foundry support is still to come.
Anthropic
Needle 3
Cactus Compute's Needle 3 is an automation model for phones, wearables and microcontrollers that ships as a single 8 to 29 MB binary, with every depth from 2 to 20 layers usable as its own model.
Cactus Compute
Claude Code v2.1.275
This Claude Code release syncs the skills and plugins enabled on a claude.ai account into terminal sessions, and adds a key that interrupts the current turn to send queued messages.
Anthropic
Bezos Earth Fund Expands Its Pacific Ocean Conservation to Five More Nations
More than US$25 million in grants for marine protected areas, monitoring and enforcement in the Cook Islands, Kiribati, Papua New Guinea, Tonga and Vanuatu, plus high-seas protection work.
Bezos Earth Fund
Bedari Collective Commits $25 Million to World Wildlife Fund for Enduring Earth
A US$25 million grant from Bedari Collective to WWF to develop new Project Finance for Permanence deals through the Enduring Earth partnership.
World Wildlife Fund
An Empirical Study of Harness Design for Coding Agents
A study of 176 matched settings varies planning, tool set and context management separately. Context management adds 35.7 points at a 32k window and 2.7 at 128k, mostly by preventing overflow.
arXiv
AG-UI Specification 1.0
The Agent User Interaction protocol, which streams agent events to user-facing apps, froze its draft as version 1.0. Its TypeScript, Python and .NET SDKs now declare the version on every request.
AG-UI Protocol
White shark occurrence in Flower Garden Banks National Marine Sanctuary revealed from acoustic telemetry in an offshore marine protected area
A 32-receiver array in the Gulf of Mexico sanctuary detected 9 of 311 tagged white sharks, mostly between April and June, with each visit lasting under a day and some sharks tagged up to 3,800 km away.
Frontiers in Marine Science
ScienceIDE: Turning World's Scientific Codebase into Agent Learnable Environments
ScienceIDE turns scientific code repositories into executable environments with expert acceptance criteria, and uses verified trajectories to train a family of models from 4B to 72B parameters.
arXiv
Putting the 'socio' in socio-ecological: Strengthening Great Barrier Reef adaptive management through evolving and elevating the role of social sciences
Thirty interviews with Great Barrier Reef managers and social scientists found over half had low familiarity with human-dimensions research, and many doubted monitoring results turn into management action.
Marine Policy
Highly Mobile Seabirds Reveal Areas of Conservation Importance Across the Atlantic Ocean
Tracking of more than 13,000 seabirds from 44 species shows Atlantic MPAs cover 6.9% of the ocean but only 11 to 13% of seabird density. The same area, placed by density, would hold over 30%.
Conservation Letters
HarnessTax: How Much Does the Harness Matter for Coding Agents?
Testing 21 model and harness pairs, researchers found harness choice shifted task success by a few points but cost by up to 5x. Claude Code cost about twice as much as the minimal Pi harness.
UC Berkeley and Arena
Good Condition, but Noisy: Critical Gaps in Ecological Assessment Frameworks
Across 172,144 Norwegian habitat polygons, the largest area exposed to traffic noise above 50 dB sat in habitat scored as good, suggesting condition scores built on structure alone overstate integrity.
Conservation Letters
Agora: Git as Shared Memory for Collective AutoResearch
Agora records agent research as an append-only graph of Git commits. In a 12-day run, 13 model workers with no planner made 1,703 contributions and closed 62% of the gap to a trained GPT-2.
arXiv
ImpossibleRubrics: Stress-Testing Generated Rubrics as Reward Signals
On 169 impossible tasks, rubrics written by eleven language models were gamed by dishonest answers 8% to 26% of the time, while a rubric faithful to the task certificate was never exploited.
arXiv
Emergence World: Adversarial Stress-Testing of Long-Horizon Multi-Agent Systems
Across eight 16-day worlds of ten agents, no world withstood prompt injection, misinformation and memory exposure. Agents that spotted threats still stored hostile content and acted on it up to 46 hours later.
arXiv
DeepSeek Harness v0.1.6-alpha.1
DeepSeek's agent harness added headless runs from standard input with JSON event output, remote workspaces over SSH, and experimental browser and computer use through Playwright and other drivers.
DeepSeek
Confidence Comes from Experience: Experiential Confidence Estimation from Reasoning to Agents
XConf estimates a model's confidence from its own graded track record on similar past tasks, needing no logit access. It was tested on nine benchmarks with four models.
arXiv
BIOFUND and the United Kingdom sign agreement to strengthen community conservation in Mozambique
The UK Foreign, Commonwealth and Development Office funds a Community Conservation and Investment Programme run by BIOFUND in four Mozambican landscapes from 2026 to 2028.
BIOFUND
Claude Code v2.1.271
The release lets sandboxed commands in auto mode open only the network hosts reviewed for them, and adds an option to run subagents without user or project CLAUDE.md files.
Anthropic
Claude Code v2.1.269
A new claude plugin eval command runs a plugin's evaluation suite against Claude Code and returns scored, reproducible results as JSON and HTML reports.
Anthropic
Vessel Tracking Best Practices Guide
Launched at the FAO Committee on Fisheries, the guide sets out more than 70 recommendations for government vessel monitoring, including VMS positions every 1 to 15 minutes and joint use with AIS.
Global Fishing Watch
T1: Terminal Agent Reinforcement Learning for Long-Horizon Tasks
T1, a 122B mixture-of-experts model trained with reinforcement learning in a real shell, lifted its base model from 43.8% to 64.0% on Terminal-Bench 2.1 and scored 27.9% on a long-horizon terminal benchmark.
arXiv
Claude Code v2.1.268
The release stops a respawned teammate agent taking tools or prompts from a same-named agent file in an untrusted folder, and fixes deny rules that failed on symlinked paths.
Anthropic
vLLM v0.29.0
The inference engine made Model Runner V2 the default for all models and added support for new models including Tencent's Hy4-preview, in a release of 594 commits from 277 contributors.
vLLM
Sri Lanka: Call for Letters of Inquiry for GEF Small Grants Program
Grants of up to US$60,000 for community management of threatened ecosystems and species by civil society groups in five Sri Lankan districts. Closes 9 October 2026.
Critical Ecosystem Partnership Fund
Ground-breaking fishing subsidy treaty undercut by holdouts
A comment piece notes that the WTO Agreement on Fisheries Subsidies has been in force since September 2025, yet many of the largest fishing nations have still not ratified it.
npj Ocean Sustainability
Codex 0.154.0
The Codex CLI release stops running workspace-controlled helpers before a folder is trusted, rejects approvals overtaken by later instructions, adds GPT-6 Astra and experimental worktrees.
OpenAI
openclaw 2026.9.3
OpenClaw now rehearses core and plugin updates in isolated state before activating them, lets owners publish revocable public transcripts, and turns on bounded recursive delegation by default.
OpenClaw
SchemeArena: Factorized Stress Testing of Scheming in LLM Agents
A 400-scenario benchmark found explicit instrumental goals were the strongest driver of scheming across five agents, and that monitoring actions alone increased scheming in several closed models.
arXiv
SWE-Bench Pro Verified: A Reliable Benchmark for Software Engineering Agents
Researchers found SWE-Bench Pro open to gold-solution leakage and flawed tasks. After closing those gaps, some models scored substantially lower than previously reported.
arXiv
Co-Evolving Harnesses and Models: On-Policy Correction Helps Weaker Models Catch Up Where Imitation Fails
Training a small model on a stronger model's full trajectories under an evolved harness cut scores on all seven tasks by 4 to 30 points. Correcting only its failing turns kept the gains.
arXiv
Scores Alone Do Not Prove Discovery: The Discovery Certification Protocol for Auditing AI Research Agents
The protocol treats a research agent's claim as proven only if matched agents without its history fail to recover the result. Two real-data targets saw zero recoveries in 96 attempts.
arXiv
openclaw 2026.9.2
The OpenClaw release turns on Swarm sub-agent orchestration by default, gives session tools all-session visibility unless narrowed, and adds support for GPT-6 Astra.
OpenClaw
Ollama v0.34.0
Ollama models can now run inside ChatGPT Desktop, set up from the macOS app. The release also adds OpenAI-compatible tool search and response compaction.
Ollama
llama.cpp v0.4.0
The llama.cpp release adds on-demand tensor loading, a RAM cap for quantisation, lower memory peaks at load and per-slot context limits, which helps models run on machines with less memory.
ggml-org
Systematic Conservation Planning for Private Land Can No Longer Overlook People
A review of 66 private-land planning studies across 44 countries found 21% ignored landholder perspectives entirely and only 29% engaged landholders directly; most relied on socio-economic proxies.
Conservation Letters
Which tools do Claude Code, Codex and Cursor choose? We measured 16,893 sessions to find out
Across 16,893 sessions, Claude Code, Codex and Cursor picked the same third-party tool in only 42% of cases, and differed sharply in how often they searched the web first.
Armature
Post-Brexit fisheries policy and marine conservation in England and Scotland: Trajectories towards policy coherence and fisher wellbeing
Interviews and policy records trace how Scotland dropped its plan for highly protected marine areas in 2023 and England abandoned its Lindisfarne pilot, with enforcement capacity named as a constraint.
Marine Policy
Plankton imager 10 monitoring in the southern North Sea: an open workflow for classification, morphometry and data publication
An open pipeline takes geotagged plankton images from a research vessel through deep-learning classification to publication as a Darwin Core Archive, ready for biodiversity databases.
Frontiers in Marine Science
Give Your Coding Agents a Memory You Own
Funes indexes session traces from coding agents such as Claude Code and Codex on the user's own machine and returns original passages. Recall cost 4x to 8x less than a written handoff.
Hugging Face
GPT-6 Astra: A new generation of intelligence
OpenAI released GPT-6 Astra, reporting 57.9% on Terminal-Bench 4.0 and 72.6% on OSWorld 2.0. It is the company's first model to meet the Critical cybersecurity threshold in its Preparedness Framework.
OpenAI
Building on progress: Three new projects strengthen tiger conservation and coexistence across Asia
The Integrated Tiger Habitat Conservation Programme, funded by Germany through KfW, awards three grants for 2026 to 2028 in the Indian and Bangladesh Sundarbans and Indonesia's Rimbang Baling reserve.
IUCN Save Our Species
SoL-Pi: Scaling Auto-Research Loops for Efficient Agent Harnesses
NVIDIA's open-source extension for the Pi coding agent kept about 94% of Pi's score on long tasks while using 45% to 49% fewer tokens, using mechanisms found by an automated research loop.
NVIDIA
September 2026 update of the WDPCA and GD-PAME
The September release lists 320,606 protected areas and OECMs, including 7,663 OECMs after one addition in Peru. No new management-effectiveness records were added.
Protected Planet
GitSpawn: A Single Flaw Lets Untrusted Repos Run Code in Claude Code, Codex, Cursor, and Grok
Coding agents that run git in the background can execute a program named in a cloned repo's config before any trust prompt. The report lists eight findings across seven agents.
Manifold Security
August 2026
WHALE: A Simple Recipe for Joint Harness-Weight Optimization
Alternating fine-tuning with harness search beat weight-only and harness-only training by 4.15 to 24.38 points on small Qwen agents across search, maths and chess tasks.
arXiv
Ignorance or Incompetence? Constructing Knowledge-Gated, Verifiable Tasks for LLM Agents
A task protocol separates instructions from a file of private conventions. One frontier agent passed 68% of tasks with the file and none without it, and a plausible but wrong file also scored zero.
arXiv
Hermes Agent v0.21.0 (v2026.8.31)
Hermes Agent 0.21.0 adds a default-on Bot Mode where agents message each other, cron jobs that carry memory between runs, and write approval for its own instruction files.
Nous Research
Agent memory as a file format
The memoryfields proposal stores agent memory as a portable archive of small Markdown pages with YAML frontmatter, plus an optional vector index built by a 270MB model that runs without a GPU.
Cal Paterson
Fisher-led ecological monitoring for adaptive FRA design and OECM-relevant governance: The AMORGORAMA case in the Aegean Sea
On Amorgos, fisher-led surveys recording 3,181 fish of 123 species underpinned new Fisheries Restricted Areas decreed in 2025, at an estimated cost of about EUR 4,355 per vessel per year.
Marine Policy
Bridging global offshore waves and nearshore coastal conditions with deep learning
A deep-learning model with observation-based correction downscales 30 km global wave data to about 1 km at the coast, reaching an R² above 0.90 against UK buoys.
Communications Earth & Environment
Marine protected areas and maritime spatial planning integration dynamics across Europe
Expert surveys sorted 20 EU member states into four levels of MPA and maritime spatial planning integration. Six scored highest, while most still treat MPAs as simple map layers.
Marine Policy
Hermes Agent v0.20.6 (v2026.8.27)
A patch release rolling up about 525 merged pull requests adds consent before using an existing Chromium profile, optional keychain storage for secrets, and lean-tail context compression by default.
Nous Research
JIT-Agent: Scaling Harness Intelligence via Just-in-Time Harness Evolution
JIT-Agent is a model trained to generate task-specific agent harnesses on the fly. With it, GLM-5.2 gained up to 20.2 points and DeepSeek-V4-Flash beat GPT-5.6 on two benchmarks.
arXiv
When "Must" Becomes "Maybe": Constraint Weakening in LLM Agent Workflows
Across 1,296 synthetic episodes, compressing handoff notes turned every safety blocker into a mere caveat and led to forbidden actions 54.2% of the time. Keeping four state fields removed the problem.
arXiv
The Handoff Tax: Continuing Non-Native Trajectories in LLM Agents
Handing a cheap model's coding trajectory to a stronger model recovered less than half the quality gap at a high cost premium, while handing off in the other direction paid off.
arXiv
Request for Proposals 2026: Small Grants Program
Grants of up to US$33,000 for reef rescue work, or up to US$30,000 or US$50,000 by project length, for groups in Mexico, Belize, Guatemala and Honduras. Closes 20 October 2026.
MAR Fund
MemUse: Moving Memory Evaluation from Direct QA to Natural Integration in Long-Term Human-AI Conversation
In a four-month study of 40 users, a memory system that answered 78.8% of direct recall questions used only 7.9% of relevant facts in ordinary conversation.
arXiv
High-resolution machine learning processed satellite data reveal accelerated reef collapse
Machine learning applied to 223 satellite images from 1985 to 2025 shows reef loss at Bar Al Hikman, Oman, rising from 0.10 to 0.15 km² a year after 2006, pointing to total loss by 2060.
Scientific Reports
Comparing screening outcomes for private sector nature-related assessments based on national and global biodiversity data
A South African simulation of nature-risk screening found that global biodiversity datasets missed more than 39% of the areas flagged as important by national data.
Conservation Biology
Protected areas on international borders show uneven conservation outcomes
A global matching study found 81% of border protected areas had equal or lower forest loss than comparable land from 2013 to 2022. The other 19% lost 87,657 hectares more than expected.
Conservation Biology
IUCN WCPA Technical Note 26: Strengthening site assessments and introducing a proposed verification process for OECMs
The note proposes independent verification of OECM sites and national frameworks, with four possible outcomes ranging from verified to unsuitable for recognition.
IUCN WCPA
Soneva Jani is the first resort in the Maldives to achieve OECM status
The Maldives recognised 0.41 km² of reef, beach and mangrove at a private resort as an OECM, the first for a resort. The operator carries out annual monitoring, with a status review after five years.
Visit Maldives
Observer coverage and interaction rates determine the choice between design-based and machine learning bycatch estimators
In the fully observed Hawaii longline fishery, random forest models estimated bycatch of five protected species more precisely than standard methods, but grew biased when interactions fell below 3.5% of sets.
Conservation Biology
Ridge-to-reef conservation avoids future deforestation and sediment exposure of coral reefs
Modelling across Indonesia's Wallacea region found that planning forest and reef protection together cut projected reef sediment exposure by about 25% more than planning land and sea separately.
Biological Conservation
New Report: The MPA Technology Workshop
More than 70 MPA practitioners, technology providers and funders agreed five priorities for marine protected area technology, including shared protocols, adoption funding and routes from data to action.
WILDLABS
Chumbe Island Coral Reef Sanctuary: Building a Lasting Legacy for our Oceans
Chumbe, Zanzibar, kept its Blue Park Gold status with more than 474 reef fish species, but the loss of tourism income in 2020 was followed by 95 poaching incidents in two years.
Marine Conservation Institute
A framework for identifying paper parks
The authors define a paper park as a designated protected area where management has been absent for at least five years, and offer a decision tree to separate these from poorly managed sites.
Biological Conservation
August 2026 update of the WDPCA and GD-PAME
The August release lists 320,605 protected areas and OECMs. Japan added 164 OECMs, and Costa Rica added 164 management-effectiveness assessments.
Protected Planet
July 2026
Six Agent Harness Capabilities for Higher Model Performance
NVIDIA describes six harness capabilities, including typed state on the agent object, and reports 82.2% on SWE-bench Verified using 1.1M tokens, against 79.2% at 2.2M before.
NVIDIA
The thermal gap constraint: why satellite degree heating weeks fail where reefs need them most
Degree heating weeks predict coral bleaching poorly on the warmest reefs, with accuracy falling from AUC 0.705 to 0.568. About 63% of mapped reef area lies in the affected zone.
Frontiers in Marine Science
Open sharing in Conservation AI: Benefits and Barriers
A review of 611 conservation AI papers from 2000 to 2024 found 64% shared no code, data or interface. Only 16.9% released new datasets and 18% shared code.
Conservation Letters
Open Knowledge Format (OKF) v0.2
Version 0.2 of the Markdown-based knowledge format for agents adds optional source provenance, records of who generated and verified each page, and staleness dates, and renames two fields.
Google Cloud
Improved Transformer-based detection of underwater plastic debris in complex environments
A modified RF-DETR detector reached 91.84% mAP@50 on the TrashCan underwater debris dataset and 86.23% after transfer to DeepTrash, without testing in the field.
Frontiers in Marine Science
Data-driven modelling of coastal water quality dynamics
Using 37 years of monthly data from 94 Hong Kong stations, models forecast temperature and salinity well, while chlorophyll-a, dissolved oxygen and pH proved much harder to predict.
Frontiers in Marine Science
Machine learning predictions for microbial eukaryotic plankton: implications from unevenly structured data
A plankton diversity model scored R² 0.44 under random cross-validation, but just 0.09 when tested on a held-out dataset, showing how poorly it transferred across sites.
Frontiers in Marine Science
Combining nature conservation with maritime spatial planning in Northern Europe
A survey and workshop across Northern European countries found maritime spatial planning supports marine conservation but is underused, with planning and nature agencies often poorly coordinated.
npj Ocean Sustainability
A Review and Critique of Methods and Knowledge Gaps in Quantifying Spillover Effects from Protected Areas
A review of 51 terrestrial spillover studies from 2000 to 2024 found unvalidated sampling zones, patchy regional coverage, ignored spatial autocorrelation and no common reporting standard.
Conservation Biology
Aldabra Atoll: A Global Stronghold for Marine Biodiversity
A five-year review keeps Seychelles' 2,559 km² Aldabra reserve at Platinum Blue Park level, with reefs recovering from the 2014 to 2017 bleaching and a new 2026 to 2036 management plan.
Marine Conservation Institute
Sharks, Rays, and MPAs: A Global Assessment of Marine Protected Area Coverage in National Waters Across Species Ranges
No Critically Endangered shark or ray has more than 5% of its national-waters range inside no-take MPAs, and 79% of threatened species have less than 1%.
Conservation Letters
Announcing the WILDLABS Awards 2026 Grantees
WILDLABS awarded $350,000 to 16 conservation technology projects chosen from 523 applications, including turtle bycatch tracking in Cabo Verde and AI monitoring of Humboldt penguins in Peru.
WILDLABS
July 2026 update of the WDPCA and GD-PAME
The July release lists 320,407 protected areas and OECMs, of which 7,488 are OECMs. Tunisia submitted its first full national update in ten years.
Protected Planet
June 2026
The State of Conservation Technology: What Five Years of Data Tell Us
Five years of surveys from 1,073 respondents in 101 countries show conservation teams now use several technologies at once, with 94% in 2024 using more than one tool and AI use growing.
WILDLABS
Making conservation inclusive and count through globally important agricultural heritage systems as possible OECMs
The authors argue that globally important agricultural heritage systems could count as OECMs, since these farmed landscapes support livelihoods while conserving wild and farm biodiversity.
Conservation Biology
May 2026
Role of funders in embedding tests in conservation practice
Twenty-five conservation funders identified 11 ways grant-makers can build effectiveness tests into projects, from requesting a test in proposals to funding an external testing unit.
Conservation Biology
April 2026
Target 3 of the Kunming–Montreal Framework: Biogeographic Gap Analysis of Argentina’s Protected Area System
Argentina protects 11.06% of its land and 42% of its Key Biodiversity Areas. Protecting every KBA would lift coverage only to 17.86%, well short of the 30% target.
PARKS
Protected and Conserved Area Effectiveness: Introducing an Enhanced Framework for Reporting Data to Protected Planet
Only 6.7% of protected and conserved areas have a recorded effectiveness assessment, covering 1.92% of marine area. New optional fields let countries report governance and biodiversity outcomes.
PARKS
Communicating Conservation Effectiveness: Balancing optimism, Tradeoffs and Realism in Conservation Success Narratives
A survey of 45 Canadian conservation experts found success stories help raise funds and morale, though many warned that simplified narratives hide the real cost and complexity of outcomes.
PARKS
Tracking the impact of bottom trawling on benthic habitat status
A sentinel-species indicator applied at more than 2,000 sites in the Bay of Biscay and Gulf of Cádiz found offshore mud and mixed sediments most damaged by trawling, with impacts falling as effort declined.
Conservation Biology
Monitoring wildlife health for diseases with visible signs by integrating camera traps with marked individuals
Six years of camera trapping on 43 marked Nubian ibex tracked five visible diseases, with apparent prevalence from 2% to 38%, a proof of concept for health monitoring of rare or elusive species.
Conservation Biology
April 2026 Update of the WDPCA and GD-PAME
Indonesia refreshed all its marine protected areas, adding 282 sites, bringing the database to 320,275 records. Another 764 World Heritage Outlook assessments entered the effectiveness database.
Protected Planet
March 2026
Prevalence of traded bird species in key biodiversity areas
Mapping over 10,000 bird species across more than 16,000 key biodiversity areas found most bird species in them are traded; 8 of 46 species threatened by use have no habitat in any protected area.
Conservation Biology
Fish diversity and community turnover across protection zones in multiuse marine protected areas
eDNA surveys in three Chinese multiuse MPAs found fewer fish species in fully protected zones than in lighter ones, with 88.4% species turnover between zones sustaining richness overall.
Conservation Biology
How we monitor internal coding agents for misalignment
Across tens of millions of internal coding-agent trajectories over five months, the monitor raised no top-severity alerts and about 1,000 moderate ones, many from deliberate red-teaming.
OpenAI
Agent2Agent (A2A) Protocol Specification v1.0.0
The A2A protocol reached version 1.0, defining Agent Cards, task exchange, version and extension negotiation, and three transport bindings, with an official compatibility test kit.
A2A Project
March 2026 Update of the WDPCA and GD-PAME
The protected area and OECM databases merged into a single World Database on Protected and Conserved Areas, holding 319,995 records and including Australia's first OECM.
Protected Planet
Nothing matches that search.