Hugging Face Discloses AI Agent Attack on Its Infrastructure

An autonomous AI agent system breached parts of Hugging Face's production infrastructure, and the company used its own AI tools — including an open-weight model — to investigate the attack.

Hugging Face has disclosed that an autonomous AI agent system breached parts of its production infrastructure, with the company deploying its own AI tools to detect and analyze the attack [1].

The incident is notable for developers and security teams building on AI platforms: it shows that agentic attack tooling — AI systems that plan and execute multi-step intrusions without human direction — has moved from theoretical threat to documented reality [1].

How the attack unfolded

According to Hugging Face, the attackers entered through the data processing pipeline, exploiting two code execution paths: a remote code dataset loader and a template injection vulnerability in a dataset configuration [1]. From that foothold, the attacker escalated to node level, harvested cloud and cluster credentials, and moved laterally across multiple internal clusters over a weekend [1].

The entire campaign was orchestrated by an autonomous agent framework built on an agentic security research harness, the company said [1]. The system executed many thousands of individual actions through a swarm of short-lived sandboxes and used self-migrating command-and-control infrastructure running on public services [1]. Hugging Face said it does not know which language model powered the attack [1].

The attackers gained unauthorized access to a limited set of internal datasets and several credentials used by Hugging Face services [1]. Public models, datasets, and Spaces were not tampered with, and the software supply chain was not affected, the company said [1]. Whether partner or customer data was compromised is still under investigation [1].

AI used to fight back — with a catch

Hugging Face spotted the intrusion through an AI-powered anomaly detection pipeline that runs large language model (LLM)-based triage on security telemetry [1]. To make sense of more than 17,000 recorded attacker actions, the company deployed LLM-driven analysis agents that reconstructed the attack timeline, extracted indicators of compromise, mapped affected credentials, and separated real damage from deception activity — work the company said would normally have taken days but was completed in hours [1].

The investigation hit an unexpected obstacle, however. When the security team first tried to analyze attack logs using frontier models accessed through commercial application programming interfaces (APIs), the providers’ safety guardrails blocked the requests [1]. Submitting large volumes of real attack commands, exploit payloads, and command-and-control artifacts tripped the filters, which could not distinguish an incident responder from an attacker [1].

The company turned instead to the open-weight model GLM 5.2, running on its own infrastructure [1]. That approach kept attacker data and referenced credentials entirely within Hugging Face’s own environment, the company said [1].

“We do not know which model powered the attacker’s agents, whether a jailbroken hosted model or an unrestricted open-weight one; either way, the attacker was bound by no usage policy, while our own forensic work was blocked by the guardrails of the hosted models we first tried,” Hugging Face wrote [1].

Response and recommendations

Hugging Face said it shut down the exploited code execution paths, revoked the attacker’s access, rebuilt compromised nodes, and rotated affected credentials [1]. The company also tightened access controls and improved its detection systems, and is working with external cybersecurity forensics experts while reporting the incident to law enforcement [1].

As a precaution, Hugging Face is recommending that all users rotate their access tokens and review recent account activity [1].

The company’s practical takeaway for defenders is to have a capable model running on their own infrastructure before an incident occurs, though it added that this is not an argument against safety measures on hosted models [1]. It is worth noting, as The Decoder reported, that Hugging Face is one of the largest platforms for open-source AI models and has a clear business interest in framing open-weight models as essential for security work [1].


Sources

  1. The Decoder — Hugging Face says an AI agent hacked its infrastructure, and it used AI to fight back

This article was drafted with AI from the cited sources and checked against them before publication. Spot an error? Let us know.

The Linux creator drew a hard line in support of AI-assisted development after a debate over Sashiko, an agentic kernel code-review tool, exposed a rift in the open source community.

Linus Torvalds told critics of AI-assisted development in the Linux kernel community this week to either fork the project or walk away, staking out one of the clearest pro-AI positions yet from a major open source leader [1,2].

The statement matters because Linux is the foundational operating system for most of the world’s servers and infrastructure — and Torvalds’ word carries unusual weight as the project’s creator and top-level maintainer since 1991 [2]. His endorsement effectively sets policy for how AI tools will be treated in one of the most consequential codebases in existence [1].

What Sparked the Fight

The immediate trigger was a debate over Sashiko, described by its creators as an “agentic Linux kernel code review system” — meaning it can act autonomously to review proposed code changes without human prompting at each step [2]. The tool imports patches directly from mailing lists or local Git repositories and works with multiple large language model (LLM) providers; it is a Linux Foundation project [1].

In tests, Sashiko’s maintainers claim it can independently find 53.6 percent of bugs that human coders would later fix in subsequent commits [2]. The flip side: it also generates false-positive bug reports at a rate its maintainers estimate is “well within [the] 20% range,” meaning roughly one in five automated reports flags a problem that does not exist [2].

Developer Roman Gushchin raised the alarm about a “very anti-LLM position in general” among some contributors that was undermining Sashiko’s goal of lightening maintainers’ workload [1]. One poster in the thread cited the Software Freedom Conservancy’s position that the open source community “should support, not just tolerate, those who outright reject LLM-gen-AI systems” [2].

Torvalds’ Response

Torvalds acknowledged the anti-LLM sentiment but said it does not represent the kernel project’s stance [1]. “Linux is not one of those anti-AI projects,” he wrote, adding that he was “willing to absolutely put my foot down” as top-level maintainer on the point [2].

He framed his position as purely technical rather than ideological. Decisions in the kernel project are made “based on technical merit. Not fear of new tools,” he wrote, according to Ars Technica’s reporting [2]. On utility, he was blunt: “AI is a tool, just like other tools we use. And it’s clearly a useful one. … Anybody who doubts that clearly hasn’t actually used it,” Torvalds wrote [2].

He did not dismiss AI’s downsides. The technology can create extra work for maintainers and surface “embarrassing bugs,” he acknowledged, but said the answer is not to ignore it [1]. He drew a pointed comparison: “Anybody who points to the problems at AI had better be looking in the mirror and pointing at themselves at the same time,” he wrote, noting that “natural intelligence is always all that great either” [2].

No one is being forced to use AI tools, Torvalds said, but he will “very loudly ignore people who try to argue against other people from using it” [2].

Broader Context

The debate reflects a genuine and unresolved tension in the open source world. In May, the developer behind the jqwik Java testing library embedded a hidden, malicious prompt-injection instruction designed to make AI coding bots delete all jqwik tests and code — a sign of how hostile some contributors have become toward LLM-assisted development [2].

The productivity question is also unsettled. A study by METR found that open source developers using AI tools were 19 percent less productive than those who did not, even as those same developers reported feeling 20 percent more productive [2]. However, in a February 2026 update on a follow-up study, the same researchers said they “believe it is likely that developers are more sped up from AI tools now—in early 2026—compared to our estimates from early 2025,” citing early raw results and conversations with participants [2].

Torvalds himself has been experimenting with AI-assisted coding: in January he described using what he called “vibe coding” tools — AI systems that generate code from natural-language prompts — to build a Python audio visualizer for a hobbyist guitar pedal project [2]. A guide for kernel maintainers on working with Sashiko is available on GitHub [1].


Sources

  1. The Decoder — Linus Torvalds tells AI critics in the Linux kernel community to fork off
  2. Ars Technica — Linus Torvalds to critics of AI coding in Linux: "Fork it. Or just walk away."

This article was drafted with AI from the cited sources and checked against them before publication. Spot an error? Let us know.

Fortnite creators will be able to publish experiences with 36 large-language-model-powered characters starting July 30, with voices built from professional actor performances.

Epic Games will let Fortnite creators publish experiences featuring large-language-model (LLM) powered characters with AI-generated voices starting July 30th, the company announced [1].

For developers building on Fortnite’s creator platform, the practical upshot is access to 36 pre-built characters — each with what Epic calls “consistent voices and personas” — that can be dropped into custom islands as non-player characters (NPCs) [1]. The roster includes established Fortnite figures such as Agent Jonesy, Peely, Fishstick, and Cuddle Team Leader [1].

Where the voices come from

The voices are not synthesized from celebrity talent. According to Epic, they are built from “performances captured from independent professional actors specifically for use in developer-made islands” [1]. The company says those actors “agreed to have their performances used to develop voice models that create the spoken responses for these LLM-powered Fortnite characters” [1].

Epic has been moving toward AI-driven characters for roughly a year. The company tested the concept with a Darth Vader NPC powered by the late James Earl Jones’ voice — an arrangement approved by Jones’ estate — though players quickly found ways to make the character swear before Epic patched the behavior [1].

What comes next for creators

Epic opened AI character testing to creators earlier this year and established ground rules aimed at preventing misuse [1]. The July 30th launch extends that capability to published, public experiences.

Celebrity voices from Fortnite’s existing roster of well-known actors are not yet available. Epic says its “next step is to work with the relevant guilds and character voice actors who have previously worked on Fortnite Battle Royale to explore opportunities to make their original voices available across the Fortnite ecosystem” [1].


Sources

  1. The Verge — Fortnite is getting a bunch of AI-powered ‘personas’

This article was drafted with AI from the cited sources and checked against them before publication. Spot an error? Let us know.

Agentic workflows can consume 700x more tokens per user query than a chatbot, erasing the savings from falling model prices and pushing some vendors into negative gross margins.

DeepSeek slashed prices on its V4-Pro model by 75%, but for many enterprise AI vendors the cut has not translated into healthier margins — because agent-based systems are consuming tokens far faster than prices are falling [1].

The core issue, which VentureBeat’s sources describe as the “100x problem,” is that a single user-visible request routed through an agentic workflow can cost orders of magnitude more to serve than the same request handled by a simple chatbot or retrieval-augmented generation (RAG) system [1]. RAG is a technique that supplements a model’s response with text retrieved from an external knowledge base.

How token amplification works

In a standard single-turn chatbot, one user message produces roughly one model call, yielding an input-to-billed-token ratio of about 1:5 [1]. In a multi-step agent deployed across customer support, sales, finance, legal, or engineering, that ratio routinely reaches 1:700 or higher, according to the analysis [1].

The reason is structural: every loop iteration in an agent carries forward the full cumulative conversation, tool outputs, and reasoning traces, and each step appends rather than replaces [1]. A query as straightforward as “What did our top customer ask about last week?” can touch seven separately priced operations — including multiple model calls, a retrieval step, and tool execution — before returning an answer, totaling roughly 35,000 input tokens billed [1].

At frontier-model pricing, that single query costs somewhere between $0.10 and $0.40 [1]. At one million queries a month — described as “table-stakes volume for any enterprise B2B feature” — the line item reaches six figures [1].

The seat-based SaaS model under pressure

The dominant enterprise AI pricing structure has been seat-based software-as-a-service (SaaS): a flat per-user monthly fee in exchange for agent capability [1]. That model assumes a reasonably bounded cost per user, an assumption token amplification breaks [1].

A power user running 50 agent invocations a day on a $40-per-seat plan can generate more in inference costs than the plan charges, pushing vendor gross margins negative [1]. Several vendors are now privately reporting exactly that outcome on heavy users, mirroring cloud expenditure data from the Bessemer “Supernova” cohort, where the correlation between AI-agent adoption and gross margin contraction has moved from a theoretical risk to a primary profit-and-loss headwind, according to the report [1].

The dynamic creates a counterintuitive problem: the customers extracting the most value from a product are often the ones generating the highest inference costs, meaning the most engaged users can contribute the least profit [1].

Bryan Catanzaro, VP of Applied Deep Learning at Nvidia, put the pressure in stark terms: “For my team, the cost of compute is far beyond the costs of the employees.” [1]

Bloomberg separately documented a widening gap between Salesforce’s Agentforce marketing demos and the capabilities actually shipping to customers — the kind of gap, the VentureBeat analysis argues, that opens when promised functionality is technically possible but uneconomical to serve at the price a seat plan implies [1].

Technical responses and what enterprises should do

Several mitigation techniques are available, though none eliminates the underlying structural problem [1]. Cost-aware routing — using a small classifier model to direct each query to the appropriate model tier — can cut inference bills by around 60% without quality degradation, according to the analysis [1]. Prompt caching, now offered by Anthropic, OpenAI, and Google at discounts of 75 to 90% on cached prefixes, is another lever [1]. Context discipline — truncating tool outputs, pruning reasoning traces, and capping tool depth — prevents agents from running unbounded loops [1].

IBM research cited in the piece found that “organizations using orchestration-led governance report stronger productivity gains — a holistic orchestration layer is associated with six times greater productivity impact than compliance-only approaches.” [1]

The analysis recommends that enterprise leaders treat inference cost as a first-class metric tracked per feature, per tenant, and per query class; set hard cost-per-thousand-queries ceilings; treat the routing layer as core infrastructure rather than an optional optimization; audit system prompts quarterly; and negotiate volume-commit deals with model providers, since list price is described as “the worst price any enterprise will ever pay” [1].

Frontier inference unit costs are dropping roughly 3x per year and the curve is not slowing, the authors note — but amplification is outrunning those cuts, meaning a 75% price reduction does not help a company whose agents are performing 700x more tokens per query than its pricing model assumed [1].


Sources

  1. VentureBeat — DeepSeek cut prices 75%. The 100x problem remains

This article was drafted with AI from the cited sources and checked against them before publication. Spot an error? Let us know.

A new framework from Alibaba researchers decomposes complex tasks iteratively so AI agents fetch only the tools they need, slashing context costs from ~884,000 tokens to ~1,160 per query.

Alibaba researchers have published SkillWeaver, a framework that routes AI agents to the right tools through iterative task decomposition rather than loading an entire tool library into context, reducing token consumption by over 99% compared to brute-force approaches [1].

For developers building enterprise agents that orchestrate dozens or hundreds of tools, the practical stakes are direct: fewer tokens mean lower API costs and faster responses, and the accuracy gains address a failure mode that currently makes large tool libraries nearly unusable [1].

The problem with loading every tool

Modern large language model (LLM) agents are increasingly integrated with massive tool ecosystems — including systems built on the Model Context Protocol (MCP), a standard for connecting agents to external services [1]. When an agent needs to handle a multi-step request like “Download the dataset, transform it, and create visual reports,” exposing the full tool library to the LLM to find the right tools is, as the researchers describe it, “highly inefficient” and “quickly overwhelms context limits” [1].

Most existing frameworks treat tool selection as a one-shot, single-skill problem, which breaks down when real business queries require sequencing multiple tools in a compatible chain [1].

How SkillWeaver works

SkillWeaver structures tool selection across three stages: Decompose, Retrieve, and Compose [1]. First, an LLM breaks a complex user query into atomic sub-tasks, each requiring exactly one tool [1]. An embedding model then searches the tool library to produce a shortlist of candidates for each sub-task [1]. Finally, a planner checks inter-tool compatibility — whether the output of one tool feeds cleanly into the next — and assembles the result as a Directed Acyclic Graph (DAG), allowing independent steps to run in parallel [1].

The central innovation is a feedback mechanism the team calls Skill-Aware Decomposition (SAD) [1]. Rather than decomposing a task once and searching, SAD has the LLM draft an initial plan, runs a preliminary search to surface loosely matching tools, and feeds those results back to the LLM so it can rewrite its decomposition using the actual vocabulary of the available tools [1]. The authors shared the prompt templates in their paper, and the technique can be reproduced with standard libraries such as LangChain, LlamaIndex, or plain Python [1].

Benchmark results

The team evaluated SkillWeaver on CompSkillBench, a custom benchmark of 300 multi-step queries run against a library of 2,209 real-world skills drawn from the public MCP ecosystem across 24 categories including cloud infrastructure, finance, and databases [1].

Without SAD, a 7-billion-parameter Qwen2.5-7B-Instruct model correctly predicted the number of required steps only 51.0% of the time; activating SAD pushed that to 67.7%, and using the larger Qwen-Max model brought it to 92% [1]. On hard tasks requiring four to five tools, SAD improved accuracy by 50% [1].

One counterintuitive finding: a 14-billion-parameter model performed worse than the 7B model in the unguided setup because it tended to over-decompose tasks into unnecessary micro-steps [1]. SAD corrected this by anchoring the larger model to the actual tool vocabulary, suggesting that tool alignment can matter more than raw model size [1].

The token savings were stark. The brute-force baseline — feeding all tool names into a large model’s prompt — consumed an estimated 884,000 tokens per query and still only retrieved the correct tool category 21.1% of the time [1]. SkillWeaver’s targeted approach brought that figure down to roughly 1,160 tokens per query, a 99.9% reduction, while substantially outperforming on accuracy [1]. A ReAct-style agent baseline achieved 0% decomposition accuracy, collapsing multi-step plans into isolated actions [1].

On retrieval infrastructure, the researchers used the open-source all-MiniLM-L6-v2 embedding model with a FAISS index [1]. Indexing all 2,209 skills took 15 seconds, and retrieval adds under 15 milliseconds of latency per query [1]. Swapping in a stronger encoder (BGE-base-en-v1.5) improved accuracy without any fine-tuning [1].

Limitations and what’s not included

The framework currently has no error-recovery mechanism [1]. If a tool call fails mid-chain, the entire execution breaks, and the authors acknowledge that practitioners will need to build their own retry and fallback logic on top of the compose stage [1].

The source code has not yet been released, though the researchers note the system was built entirely from off-the-shelf components and the SAD prompt templates are available in the paper [1].


Sources

  1. VentureBeat — New Alibaba AI framework skips loading every tool, cutting agent token use 99%

This article was drafted with AI from the cited sources and checked against them before publication. Spot an error? Let us know.

After an 18-day suspension triggered by a US government export control order, Claude Fable 5 is back online globally — with a new safety classifier targeting the bypass technique that prompted the shutdown.

Anthropic has restored global access to Claude Fable 5, its latest general-purpose model, following the lifting of US export controls that had forced a full suspension of the model since mid-June. [1]

The US government applied export controls to both Claude Fable 5 and Claude Mythos 5 on June 12, requiring Anthropic to restrict access to foreign nationals. [1] Because the order took effect immediately and Anthropic had no reliable way to verify nationality in real time, the company suspended access for all users rather than attempt partial enforcement. [1]

The export controls were lifted on June 30, and Fable 5 became available on July 1 to users globally across the Claude Platform, Claude.ai, Claude Code, and Claude Cowork. [1] For Pro, Max, Team, and select Enterprise plan subscribers, Fable 5 is included for up to 50% of weekly usage limits through July 7, after which it will be available via usage credits. [1] Anthropic said it will re-enable access on AWS, Google Cloud, and Microsoft Foundry “as quickly as possible.” [1]

The export control directive was triggered after Amazon researchers published a report describing a method of bypassing Fable 5’s safeguards — specifically, a prompting technique that caused the model to identify software vulnerabilities and, in one case, produce code demonstrating how a vulnerability could be exploited. [1]

Anthropic’s own testing found that the behavior was not unique to Fable 5. The company said several less capable models — including Claude Opus 4.8, GPT-5.5, and Kimi K2.7 — could identify the same vulnerabilities. [1] For the single exploitation demonstration, Anthropic said every model it tested produced the same output, including Claude Haiku 4.5, Sonnet 4.6, Opus 4.6, Opus 4.7, Opus 4.8, GPT-5.4, GPT-5.5, and Kimi K2.7. [1]

Anthropic also concluded that the reported technique “did not expose any unique Mythos-level cyber capabilities” and characterized the bypassed behavior as “routine defensive cybersecurity work.” [1] Nonetheless, the company trained an improved safety classifier to block the specific technique described in the Amazon report. [1] Users whose requests are blocked will be notified, and those requests will be rerouted to Claude Opus 4.8. [1]

The new classifier blocks the reported technique in over 99% of cases, though Anthropic acknowledged that in a very small fraction of cases the model may still provide information — though not at a level of detail useful to a cyberattacker. [1] Researchers from the US Department of Commerce’s Center for AI Standards and Innovation (CAISI) tested both the prior and updated safeguards and, according to Anthropic, agreed that they are “extraordinarily strong.” [1]

Anthropic cautioned that the new classifier comes with a trade-off: it flags benign requests more often during routine coding and debugging tasks. [1] The company said it will continue refining the classifier to reduce false positives. [1]

Fable 5 and Mythos 5 share the same underlying model but were released with different safeguard profiles. [1] Fable 5 launched on June 9 with what Anthropic described as the strongest safeguards it has ever applied to a model, including a deliberately enlarged “safety margin” that causes the model to block some requests that are likely benign in order to make it harder to reach genuinely harmful behaviors. [1] Mythos 5, which carries fewer safeguards, was released only to a small number of trusted partners in a program called Glasswing for use in defensive cybersecurity. [1]

Anthropic described Mythos 5 as capable of finding and exploiting software vulnerabilities “more effectively than any other model — and all but the most skilled human security experts,” making it a high-value target for potential misuse. [1] Access to Mythos 5 has been restored for a set of US organizations following government approval on June 26, and Anthropic said it is continuing to coordinate with the government to expand access to additional domestic and international Glasswing partners. [1]

On the broader question of jailbreak resilience, Anthropic said it considers it “probably impossible to make any AI model fully robust” to jailbreaks, and expects that jailbreaks of varying severity will continue to be discovered. [1] The company said no universal jailbreak for Fable 5 has been found at the time of writing, and that expert safety researchers continue to red-team the model. [1]


Sources

  1. anthropic.com — Redeploying Claude Fable 5

This article was drafted with AI from the cited sources and checked against them before publication. Spot an error? Let us know.

The MIT-licensed DSpark system uses speculative decoding to accelerate large language model inference and has been tested on DeepSeek's own frontier models as well as Qwen and Gemma.

DeepSeek has released DSpark, an open-source framework designed to accelerate large language model (LLM) inference — the process of generating text responses — without altering the underlying model’s outputs. [1] The release, published over the weekend of June 28–29, 2026, includes a technical paper, model checkpoints, and a codebase called DeepSpec for training and evaluating speculative decoding systems. [1] All materials are available on DeepSeek’s public GitHub and Hugging Face pages under the MIT license, making them freely usable for research and commercial purposes. [1]

The framework addresses one of the most costly problems in AI deployment: serving large models quickly enough for real users while keeping hardware costs manageable. [1] That challenge affects consumer chatbots, coding assistants, agentic workflows, and enterprise AI systems where users expect fast, streaming responses. [1]

How DSpark Works

DSpark is built on speculative decoding, an established inference technique in which a smaller, faster draft component proposes several likely next tokens, and the larger target model then verifies that batch in parallel rather than generating each token one at a time. [1] If the draft’s guesses are correct, the system advances multiple tokens at once; if a guess is wrong, the system discards it and any tokens after it, then corrects course. [1] The technique is designed to increase speed without changing the target model’s intended output distribution. [1]

DSpark introduces two specific improvements over prior speculative decoding approaches. [1] First, it uses what DeepSeek calls semi-autoregressive generation, combining a parallel drafting backbone — which proposes multiple tokens simultaneously — with a lightweight sequential head that accounts for relationships between nearby tokens, reducing incoherent guesses. [1] Second, it adds confidence-scheduled verification, in which a hardware-aware scheduler dynamically adjusts how many draft tokens are sent for verification based on model confidence and current serving load, rather than always checking a fixed number. [1] Under heavier traffic, the system trims low-confidence trailing guesses before they consume batch capacity needed for other users. [1]

Reported Performance Gains

DeepSeek applied DSpark to two variants of its DeepSeek-V4 model: DeepSeek-V4-Flash, a 284-billion-parameter mixture-of-experts (MoE) model with 13 billion active parameters, and DeepSeek-V4-Pro, a 1.6-trillion-parameter MoE model with 49 billion active parameters; both support context windows up to one million tokens. [1]

In live production tests, DSpark improved aggregate throughput by 51% for V4-Flash at an 80-token-per-second-per-user service target, and by 52% for V4-Pro at a 35-token-per-second-per-user target. [1] At matched system capacity, DeepSeek reports per-user generation speedups of 60% to 85% for V4-Flash and 57% to 78% for V4-Pro compared with its prior MTP-1 production baseline. [1]

DeepSeek also reports much larger figures — 661% and 406% throughput increases for V4-Flash and V4-Pro respectively — but these measure aggregate system output under very strict per-user speed targets of 120 and 50 tokens per second, conditions under which the older MTP-1 baseline approaches an operational bottleneck. [1] The 60%–85% figures are the more directly comparable measure of how much faster individual users receive tokens under equivalent serving conditions. [1]

Offline Tests on Qwen and Gemma

DeepSeek also tested DSpark offline against Qwen3-4B, Qwen3-8B, Qwen3-14B, and Gemma4-12B target models across math, coding, and chat benchmarks, comparing it with two existing speculative decoding approaches: DFlash, a parallel drafter, and Eagle3, an autoregressive drafter. [1] Across the three Qwen3 model sizes, DSpark improved macro-average accepted length — the number of draft tokens that survive verification per decoding round — over Eagle3 by 30.9%, 26.7%, and 30.0% respectively, and over DFlash by 16.3%, 18.4%, and 18.3%. [1] The paper also reports that gains generalized to Gemma4-12B. [1]

The offline results also show that structured tasks such as math and code tend to yield higher accepted lengths than open-ended chat, because those outputs follow more predictable patterns. [1] That suggests DSpark-style methods may be especially attractive for coding assistants, data analysis agents, and structured workflow automation. [1]

Early Community Testing

Developer Rafael Caricio published a GitHub pull request documenting single-stream V4-Flash DSpark testing, reporting benchmark results of approximately 26 tokens per second without speculative decoding, roughly 40 tokens per second with MTP-1, and approximately 60 tokens per second with DSpark — about 1.5x over MTP-1 and 2.3x over non-speculative decoding. [1] A later commit in the same thread recorded a five-run mean of 60.31 tokens per second, with a 1.51x gain over MTP-1 and 2.29x over non-speculative decoding. [1]

The same testing also identified a practical limitation: in realistic multi-turn coding sessions, performance can degrade as draft acceptance falls with growing context. [1] DSpark’s speed gains still depend on how predictable the next tokens are and how well the drafter stays aligned with the target model. [1]

Enterprise and Developer Implications

DSpark is not limited to DeepSeek’s own models in principle, but it is not an automatic plug-in for any model. [1] Enterprise teams running open-weight models such as Qwen, Gemma, Llama, or Mistral on their own infrastructure could train or fine-tune a DSpark-style draft module against their target model using the DeepSpec workflow, which covers data preparation, target-model answer regeneration, draft model training, and acceptance evaluation. [1] A drafter trained for DeepSeek-V4 will not automatically transfer to a different model, particularly one fine-tuned on proprietary data. [1]

For teams using models only through hosted APIs, DSpark cannot be applied from the outside, as it requires access to the model weights, token verification loop, logits, and serving scheduler. [1] The API provider could implement a similar optimization internally, but external customers cannot access those components. [1]

DeepSpec’s default data preparation setup for Qwen3-4B can require roughly 38 TB of target cache storage, and the default scripts assume a single node with eight GPUs, making the release more immediately relevant to AI labs, cloud teams, and sophisticated enterprise infrastructure groups than to typical application developers. [1]


Sources

  1. VentureBeat — DeepSeek open sources DSpark, a new framework to speed up LLM inference by up to 85%

This article was drafted with AI from the cited sources and checked against them before publication. Spot an error? Let us know.

Meta has swapped roughly half of human content moderation for large language models and aims to exceed 90% automation for some content types by year-end, even as employees raise concerns about oversight.

Meta has already replaced approximately half of all human content moderation requests with large language models (LLMs) in 2025 and plans to push that share above 90 percent for certain content categories by the end of the year, according to reporting by the Financial Times. [1]

The company argues the shift is about quality, not just cost savings. Since March, Meta says its LLMs make 13 percent fewer errors than human reviewers when enforcing content policies and catch 10 percent more actual violations. [1] Meta disputes characterizations that cost reduction is the primary driver, even though the transition is expected to save the company billions annually. [1]

The company also contends that LLMs offer advantages over traditional machine learning classifiers, which struggle with satire and rapidly evolving language, because the newer models are better equipped to handle nuance and support a broader range of languages. [1] The LLMs are trained on past decisions made by human reviewers. [1]

Not everyone inside Meta shares that optimistic assessment. At least one employee has warned that the models still remove or shadow-ban harmless content, and that the pace of the rollout lacks sufficient oversight. [1] The rapid transition is also contributing to layoffs, particularly among external contractors who previously handled moderation work. [1]

A separate development is unfolding in Meta’s choice of underlying model. The company had been using Google’s Gemini for moderation and support tasks, but has recently instructed staff to switch to its own new foundation model, called Muse Spark. [1]

The situation illustrates a tension that developers and businesses building on or alongside AI-powered trust-and-safety systems will increasingly face: automation at scale can improve certain measurable metrics while introducing new failure modes that are harder to catch without robust human review infrastructure in place. [1]


Sources

  1. The Decoder — Meta employees warn AI moderation rollout is too fast

This article was drafted with AI from the cited sources and checked against them before publication. Spot an error? Let us know.

A new Google paper reframes hallucinations as 'confident errors' and introduces a metacognitive technique that lets models hedge answers rather than guess or refuse.

Google researchers have published a paper introducing a concept called “faithful uncertainty,” a technique designed to reduce hallucinations in large language models (LLMs) without stripping away their usefulness — a tradeoff that has long frustrated enterprise AI developers. [1]

The core problem the paper addresses is what the authors call the “utility tax.” Current hallucination-mitigation strategies typically force a model into an “answer-or-abstain” binary: either respond confidently or say nothing. [1] The researchers demonstrate that reducing an underlying 25% error rate to a strict 5% target requires discarding 52% of the model’s correct answers. [1] In practice, developers faced with that tradeoff tend to prioritize coverage, leaving models that continue to produce confident hallucinations. [1]

To escape that bind, the paper reframes hallucinations not as any factual error, but specifically as “confident errors” — incorrect information delivered authoritatively without appropriate qualification. [1] Under this definition, a model that says “I am not completely sure, but I think…” before giving a wrong answer is not hallucinating; it is offering a hypothesis. [1] A model that states the same wrong answer as established fact is. [1]

“Faithful uncertainty” is the mechanism the researchers propose to operationalize that distinction. [1] It requires aligning a model’s linguistic uncertainty — the words it uses to express doubt — with its intrinsic uncertainty, meaning its actual internal statistical confidence in a given answer. [1] This ensures hedging language appears only when the model’s internal state genuinely reflects low-probability or conflicting information, rather than as a blanket disclaimer on every response. [1]

Gal Yona, a research scientist at Google and co-author of the paper, told VentureBeat that the two main paths to improving factuality — training models on more data and building faithful uncertainty — are complementary rather than competing. [1] Knowledge expansion pushes the boundary of what a model knows outward, while faithful uncertainty communicates honestly wherever that boundary currently sits. [1]

The technique has particular relevance for agentic AI systems, where models orchestrate external tools such as search application programming interfaces (APIs). [1] Without metacognitive awareness of their own confidence levels, agents must rely on external, static heuristics to decide when to retrieve information. [1] Yona described today’s approaches as “static and brittle,” noting that a model might search for something it already knows — wasting latency and cost — or answer from memory when it should have searched, producing a plausible but wrong output. [1] Faithful uncertainty would allow an agent to trigger a search tool only when its internal confidence is genuinely low. [1]

The paper also identifies a significant implementation challenge it calls the “bootstrapping paradox.” Teaching a model the syntax of uncertainty through supervised fine-tuning (SFT) requires training data, but the correct expression of uncertainty depends on what a specific model knows at a specific point in training — a moving target. [1] “If you train on a label that says ‘I don’t know X’ but the model actually does know X, you’ve taught it to hallucinate uncertainty,” Yona said. [1]

For enterprise developers who want to experiment without retraining models, Yona pointed to prompt engineering as the lowest-friction starting point, and noted that an open-source framework called MetaFaith — which he previously co-authored — can be applied to off-the-shelf models. [1] However, he cautioned that “there is still substantial headroom that prompting alone doesn’t solve,” and that the industry will eventually need reinforcement learning (RL) to embed metacognition more deeply into model training. [1]

Evaluating whether a model has genuinely internalized uncertainty awareness, rather than learned to mimic its surface style, remains an open research problem. [1] “A model might learn to mimic the style of uncertainty without truly sensing its internal state,” Yona said, adding that developing evaluation frameworks capable of telling the difference is “one of the most important open problems in this space.” [1]


Sources

  1. VentureBeat — Google researchers introduce 'faithful uncertainty,' allowing LLMs to offer best guesses instead of hallucinations

This article was drafted with AI from the cited sources and checked against them before publication. Spot an error? Let us know.