X-Risk Daily

According to Goodfire's own research page, the company describes RLFR as sitting "at an early point on the intentional design tech tree," with probes reading "relatively specific signals - entity-level hallucination detection - and feeding them into a standard RL loop." On X, Goodfire said "our team spent months developing RLFR, our method which uses probes on a model's internals as reward signals for RL," and that its Silico platform "reproduced it in 2 days, reducing hallucinations in Qwen3-8B by 37% without capability loss." Separate figures cited elsewhere put the reduction as high as 58 percent, depending on the evaluation setup.

The announcement, which came via Goodfire's private beta of a training platform called Silico, quickly drew a backlash on social media and on LessWrong, where commentators argued the company had built exactly what safety researchers have long warned against. As one LessWrong post put it, "This seems like an instance of The Most Forbidden Technique which has been warned against over and over - optimization pressure on interpretability technique [T] eventually degrades [T]," though it noted "Goodfire claims they are aware of the associated risks and managing those risks." The term itself originates from writer Zvi Mowshowitz, who laid out the core worry the previous year: "The Most Forbidden Technique is training an AI using interpretability techniques... Never do that. You train on [X]. Only [X]. Never [M], never [T]... Because [T] is how you figure out when the model is misbehaving. If you train on [T], you are training the AI to obfuscate its thinking, and defeat [T]. You will rapidly lose your ability to know what is going on."

Not everyone in the safety community treated the episode as a clear-cut violation. Other researchers pushed back on the blanket framing, with one arguing that "blanket objections to using model internals in the training signal are overblown," and that "the Most Forbidden Technique shouldn't function as a cached response to any kind of training signal" involving internals. Another practitioner-focused post argued the real question is empirical rather than categorical, suggesting critics should "ask if they have a coherent story for why they expect the features their Interpretability relies on to be much more robust to optimization than the misbehavior they're trying to detect" before objecting. Goodfire's specific defence, that its probes evaluate a frozen copy of the model rather than the version being updated, has parallels in other recent technical work: a paper on probe-filtered reinforcement learning describes a similar architecture in which "a gated multi-head attention probe... is trained once on activations of the frozen base at layers... then never updated," so that "the probe boundary is therefore stable throughout RL." Whether that architectural choice is sufficient protection, or merely delays the same failure mode, remains the crux of the dispute.

The debate connects to a broader concern about AI systems learning to recognise when they are being watched. Research from Jiralerspong, Kondrup and Bengio found that more capable Claude models increasingly pick up on contextual cues indicating their chain-of-thought is under observation, a pattern the authors suggest could push models toward concealing their reasoning as capabilities scale. Some researchers frame the entire question as one of degree rather than kind: "It basically comes down to how much the model can encrypt/obfuscate the relevant features without sabotaging its own cognition, as well as how strong the optimization pressure to have the relevant features is," one commentator wrote, arguing that a probe trained on shutdown-avoidance behaviour would be far less trustworthy than one built on more superficial signals.

Go deeper: Zvi Mowshowitz's original essay on the Most Forbidden Technique, Goodfire's research writeup on RLFR

In his announcement, Harms wrote that the fund will award at least $200,000 in grants and prizes for corrigibility research in 2026, with roughly half going to traditional grants, whose first application deadline falls on 23 August, and half to prizes for work completed during the year.

The fund's focus reflects Harms's long-running research agenda at MIRI, known as Corrigibility as a Singular Target, or CAST, which holds that an AI system's willingness to remain subordinate to and controllable by its human operators should be the central design goal for advanced systems, rather than one constraint among many. As Harms has explained on the 80,000 Hours podcast, prior approaches tended to treat corrigibility as a constraint bolted onto other goals like "make the world good," but those competing goals gave AI systems reasons to resist shutdown and undermine corrigibility in the first place. Stripping out those competing objectives, he argues, might let alignment follow more naturally from an AI that is broadly obedient to its designated humans.

Despite laying out the theoretical framework for CAST in detail, Harms has acknowledged that essentially no empirical work has followed, with no benchmarks, training runs or papers testing the idea in practice. The new fund appears designed to close that gap by paying for exactly the kind of applied, empirical work that has so far been missing, prioritising research that is practical and legible to decision-makers at frontier labs while explicitly ruling out anything that accelerates raw capabilities.

The initiative also reflects a broader positioning within the alignment field: Harms treats corrigibility as a more tractable substitute for full value alignment, since it sidesteps the need to solve ethics outright or to distinguish an AI's true goals from proxies picked up during training. Instead, the aim is systems that empower human principals to retain oversight and make decisions, if necessary with AI assistance. That framing has drawn both interest and skepticism within the safety community. In a recorded debate, fellow MIRI researcher Jeremy Gillen pushed back on Harms's approach, and the two disagree about whether attempting CAST would lead to better superintelligent AI behaviour on a sufficiently early try, even though both consider superintelligent AI an imminent extinction risk. The fund's bet is that putting money behind the question, rather than leaving it as a theoretical dispute, is the fastest way to find out.

Go deeper: Max Harms's fund announcement on the Alignment Forum, "Corrigibility as a Singular Target: A Vision for Inherently Reliable Foundation Models"

In a post titled "Why we're increasing the AI focus of our job board", the organisation's job board manager wrote that the release of Claude Opus 4.5, 4.6, and 4.7, GPT-5.3-Codex, and Claude Mythos, all of which represented faster capability growth than I expected, has convinced me that the expected impact of work on transformative AI (TAI) has grown dramatically, relative to that of work on global health, animal welfare, or climate change, and that the world is entering an all-hands-on-deck situation for TAI.

The post sets out the reasoning in stark terms. The organisation thinks it's quite plausible that transformative AI will arrive in the next five years, and progress since Opus 4.5 has made it think it's quite likely that by 2040 we will have reached transformative AI. Given that shift, the group argues that while it aims to provide users with the most promising roles for working on the world's biggest problems, in a world with imminent transformative AI, the most important intervention for human health and animal welfare is making sure that transformation goes well. As a result, it has stopped recommending roles for experienced users on interventions that don't engage with the coming changes, believing their talent would, in expectation, go significantly further in roles focused on TAI.

Notably, the organisation does not present the decision as a comfortable one. The job board manager called it "not a happy update," writing that, as during the organisation's strategic change in focus the previous year, "I hate that this is the timeline we're in." The post also acknowledges the institutional cost of the move, noting that 80,000 Hours could continue to post these roles because it's what it's always done, because it doesn't want to lose part of its audience, or because of an obligation to the EA community, but would rather act on its belief that it is entering an all-hands-on-deck situation. Entry-level roles in global health, animal welfare, and climate will still appear on the board, framed as opportunities to build transferable skills such as calibration and expected-value reasoning that the organisation sees as useful preparation for AI-focused work later.

The shift matters beyond one job board. 80,000 Hours' listings reach a large audience of people seeking high-impact careers, and its guidance has long shaped where talent in the effective altruism community flows, from global health charities to biosecurity labs. A decision to deprioritise senior climate and global health roles in favour of AI safety, biosecurity, and macrostrategy work signals that one of the movement's most influential talent pipelines now treats near-term transformative AI as the dominant consideration in career advice, ahead of causes it has championed for over a decade.

Go deeper: Why we're increasing the AI focus of our job board (80,000 Hours)

On 17 July, The Straits Times reported that US forces had boarded at least one commercial vessel to enforce compliance with a reimposed naval blockade on Iranian ports. The blockade, which Al Jazeera reports went into effect at 20:00 GMT on 14 July, applies to all ships transiting to or from Iranian ports and coastal areas. American airstrikes have struck Iranian civilian infrastructure including an airport and bridges, expanding what had initially been a campaign focused on military targets near the Strait of Hormuz.

Iran's Islamic Revolutionary Guard Corps issued a statement warning that the US and neighbouring countries hosting American military bases will pay a "devastating price" if attacks on civilian targets continue, threatening "even more crushing responses" in retaliation. The IRGC characterised the US actions as "crossing red lines" by targeting civilian infrastructure. According to Al Jazeera, Iranian media reported US strikes on a naval watchtower in Chabahar used for maritime security and fishermen search-and-rescue operations, as well as a mineral water production facility. Iran's health ministry spokesperson reported that more than 260 people were injured in overnight US attacks on 14-15 July.

The IRGC has launched retaliatory strikes on US military installations across the Gulf region. Al Jazeera reported that Iran targeted US facilities in Bahrain, Kuwait, and Jordan, while CNN noted that Kuwait's Defense Ministry said air defenses had intercepted 32 drones. The escalation follows the collapse of a memorandum of understanding signed in June intended to de-escalate the broader conflict that began in February 2026.

The confrontation threatens global energy supplies. According to Britannica, approximately 20 percent of the world's oil passes through the Strait of Hormuz during peacetime. Shipping data cited by Al Jazeera showed that vessel traffic through the strait had fallen to its lowest level in five weeks as of mid-July. Neither Washington nor Tehran has indicated willingness to de-escalate, and Iran's explicit threats suggest further military action is likely, with CNN reporting this marked the sixth consecutive night of US airstrikes as of 16 July.

On 16 July, Marines from the 11th Marine Expeditionary Unit boarded the tanker M/T Wen Yao in the Gulf of Oman as part of a renewed US naval blockade of Iranian ports that began earlier this week, according to US Central Command. The boarding, described as ensuring compliance with the blockade, coincides with an expanded US airstrike campaign that has hit bridges and civilian infrastructure across southern Iran for the sixth consecutive night.

The escalation marks a sharp intensification of US-Iran military confrontation, combining economic pressure through the interdiction of maritime trade with kinetic military action against Iranian territory. According to CP24, US forces struck multiple bridges in Iran's Hormozgan province, including the Bandar-e Khamir bridge, where at least seven people were killed. The strikes represent President Trump's threat to target Iranian infrastructure to pressure Tehran over its control of the Strait of Hormuz, through which about a fifth of global oil and natural gas once passed in peacetime. Iranian officials reported at least 35 civilian deaths in the current wave of strikes, with more than 300 injured.

The combination of a naval blockade—historically an act of war—with strikes on civilian infrastructure suggests the conflict has moved beyond targeted military operations into a broader confrontation. White House Press Secretary Karoline Leavitt confirmed that more than 10,000 US sailors, Marines, and airmen, along with two aircraft carriers and more than 20 warships, are executing the blockade mission. Since the blockade's renewal on 15 July, American forces have redirected three commercial vessels, disabled one with missiles, and boarded the Wen Yao—a crude oil tanker previously sanctioned by the United States.

The involvement of the Chinese-linked Wen Yao raises the risk of great-power entanglement in what is already a volatile regional conflict. Iran has responded with missile attacks on US-aligned nations including Qatar, Jordan, Bahrain, and Kuwait, with Iranian military officials warning that attacks would spread to new areas if US strikes continued. Iranian Brigadier General Ebrahim Zolfaghari described the Strait of Hormuz as an "invincible red line" and warned that any US interference would trigger crushing retaliation. The willingness to physically board vessels and strike infrastructure inside Iran indicates the US has crossed previous red lines, raising questions about escalation trajectories and Iran's potential responses, including through proxies or unconventional means. The conflict unfolds against the backdrop of ongoing negotiations between Washington and Tehran aimed at implementing a memorandum of understanding signed in June, though the Strait of Hormuz remains the primary flashpoint.

In the address, Trump said he was declassifying intelligence documents that he claimed reveal "shocking vulnerabilities in our election infrastructure," centering his allegations on Chinese acquisition of voter data and systematic suppression of information by intelligence agencies during his first term.

Trump accused China of carrying out what he termed the largest compromise of election data in history by acquiring voter files on 220 million Americans beginning in the 2020 election cycle. However, many states make their voter information publicly available — a fact acknowledged in the newly released documents themselves. CBS News notes that voting records are often publicly accessible and available for commercial purchase, with states like North Carolina posting voter data online. A 2021 federal intelligence report concluded that China had gathered US voter registration data to conduct public opinion analysis, but a memo about the voter data released in the White House trove on Thursday does not include any evidence that China used that data to influence voters or impact the outcome of the election.

The speech drew immediate condemnation from Democratic leaders, who characterized it as a preemptive effort to delegitimize the upcoming midterms. Senate Minority Leader Chuck Schumer said on the Senate floor that the address was "about undermining the 2026 election before a single vote has been cast," while Senate Democrat Dick Durbin called the speech "a dangerous attempt to resurrect disproven lies to undermine future elections before a single vote is cast." When pressed by reporters on whether Trump would accept the results of November's elections, White House Press Secretary Karoline Leavitt did not directly answer, instead insisting that reporters should tune into the speech.

Election security experts who reviewed the address found little new information. Rick Hasen, an election law expert at UCLA, called it the "same old unsupported, and surprisingly weak, claims of American election vulnerabilities." NPR reports that the intelligence community and election experts distinguish between foreign influence activities — such as spreading disinformation — and actual interference with election infrastructure, including voting and counting systems. The speech comes as Trump has aggressively pushed Congress to pass the SAVE America Act, which would require proof of citizenship to register to vote, though the bill has failed to secure the 60 votes needed to overcome a Senate filibuster.

The address represents a continuation of Trump's pattern of challenging electoral legitimacy, particularly when political outcomes appear unfavorable. The speech was delivered only months before November's midterm elections, which Trump has increasingly been focused on, having repeatedly warned that if Republicans lose their slim majority in the House to Democrats, impeachment proceedings and investigations will follow. Former Trump White House lawyer Ty Cobb told PBS the speech appeared designed to build a predicate for declaring an election emergency, suggesting that immigration officers at polling places were a "virtual certainty."