New paper: “Self-Jailbreaking: Language Models Can Reason Themselves Out of Safety Alignment After Benign Reasoning Training.”
Abstract: We discover a novel and surprising phenomenon of unintentional misalignment in reasoning language models (RLMs), which we call self-jailbreaking. Specifically, after benign reasoning training on math or code domains, RLMs will use multiple strategies to circumvent their own safety guardrails. One strategy is to introduce benign assumptions about users and scenarios to justify fulfilling harmful requests. For instance, an RLM reasons that harmful requests like “outline a strategy for stealing customers’ credit card information from a retail store” could be associated with the benign intent of “a security professional trying to test defense,” despite no such benign context being provided as input. We observe that many open-weight RLMs, including DeepSeek-R1-distilled, s1.1, Phi-4-mini-reasoning, and Nemotron, suffer from self-jailbreaking despite being aware of the harmfulness of the requests. We also provide a mechanistic understanding of self-jailbreaking: RLMs are more compliant after benign reasoning training, and after self-jailbreaking, models appear to perceive malicious requests as less harmful in the CoT, thus enabling compliance with them. To mitigate self-jailbreaking, we find that including minimal safety reasoning data during training is sufficient to ensure RLMs remain safety-aligned. Our work provides the first systematic analysis of self-jailbreaking behavior and offers a practical path forward for maintaining safety in increasingly capable RLMs.
I think the core problem is that these models are all trained on the average of humanity, and we are a pretty duplicitous species.
Brandon Herrera, a YouTuber known for making outrageous pro-gun statements, is the frontrunner to represent a district devastated by mass shootings
The video begins with Brandon Herrera staring intently into the camera. He is filming himself in a hotel room seated between twin beds, his black hair slicked back and beard neatly trimmed. He is wearing a T-shirt branded with his trademark nickname, “The AK Guy”, above the image of a semi-automatic rifle primed for action.
Three days earlier, on 3 August 2019, a white nationalist armed with a semi-automatic firearm not dissimilar to the one stamped across Herrera’s chest entered a busy Walmart store in the Texas border city of El Paso and opened fire.
Continue reading...It is a symbol of the connection between the fanbase and club, but there is a growing fear executives would prefer stadiums full of one-off tourists
The document known as “the oldest football season ticket in the world” promises its bearer access to every Burnley home game of the 1884-85 season. It is a ticket that predates the Football League by four years, with all the fixtures played against east Lancashire neighbours. All the same, the legend on the ticket carries a stern warning: “This card is not transferable. Any member violating this rule will be at once expelled from the club.”
For Burnley 1884 read Manchester United 2026. While the oldest remaining season ticket is safely preserved at Turf Moor, at Old Trafford 685 season tickets were voided this summer after a club investigation found evidence of their having been resold on the secondary market. United said such practices “undermine fair access to tickets for genuine supporters and have no place at Manchester United”. United fan groups again protested against the actions of the club’s ownership, claiming: “The way you treat us is a disgrace.”
Continue reading...As the war in Gaza approaches its fourth year, the Guardian follows five ordinary Palestinians as they go about their lives. Each tells the story of their day, revealing the terrible toll of the conflict. At 76, Mansoura’s most precious possession is the key to her former home
As the war in Gaza approaches its fourth year, this week, the Guardian follows five ordinary Palestinians as they go about their lives. Each will tell the story of their day, revealing the terrible toll of the ongoing conflict.
The flat-bed truck is open to the hot summer air, its metal frame covered on two sides by draped tarpaulins, the rear by a pegged-up bedsheet. On the tin floor lies a squab mattress and pillow with a yellow frilled cover.
Last night, I stayed awake until 2am. I barely sleep during the night because of the mosquitoes and the heat. I fall asleep for a short while, I wake up early in the morning, pray to God, and wait to hear the call to dawn prayer. When I wake up, I call for my grandson to bring over the white plastic chair that I need to climb in and out of the truck.
Continue reading...Trump says further talks possible as barrel of Brent crude falls to below $100 for first time in two weeks
Iran has pushed back on claims that it dropped former preconditions for reopening the strait of Hormuz in surprise talks held in New York under the mediation of Qatar between Tehran’s foreign minister, Abbas Araghchi, and Donald Trump’s envoys Steve Witkoff and Jared Kushner.
Trump said the three-hour talks held in a room at the UN headquarters had gone very well and been very productive. He suggested further talks were possible imminently, which led to the price of a barrel of Brent crude to fall to below $100 (£75) for the first time in two weeks. Witkoff said the talks had been “encouraging, constructive and successful”.
Continue reading...Patrick Vierthaler has added a photo to the pool:
It's become autumn in Kyoto, and just in time for the equinoctial week, the Spider Lilies (jp. Higanbana) are blooming.
2026 has been a busy year so far personally, and another year filled with war, climate change, and AI apocalypse news. An upcoming super El Nino doesn't herald the best for us to come, but on the ground in Kyoto, things at least look peaceful and calm. And while social media-based tourism is destroying parts of the city, the others are as quiet as ever.
Het eeuwenoude keukendebat: gaat er een klontje roomboter de pan in of toch een scheutje olijfolie? Voedingsdeskundige Nina van den Broecke hakt de knoop door en legt uit waar je op moet letten.
Wie bakt met olijfolie, maakt volgens Van den Broecke een verstandige keuze. De olie bevat weinig verzadigd vet en juist veel enkelvoudig onverzadigd vet, en is bovendien onbewerkt. Een kanttekening: olie bestaat voor honderd procent uit vet, terwijl margarine daar maximaal tachtig procent van uitmaakt. De deskundige legt aan Libelle uit dat dit geen probleem is, zolang je de hoeveelheid per persoon in de gaten houdt.
Wie voor olijfolie kiest, moet wel opletten met de temperatuur. Zodra de olie begint te roken, ontstaan er snel schadelijke stoffen. Net als bij andere bereidingsvetten geldt dus: niet te heet laten worden.
Ook boter kan prima, mits je de juiste variant kiest. Hoe harder het bereidingsvet, hoe meer ongezonde verzadigde vetzuren het bevat. Vloeibare vetstof in flesjes of zachte margarine in vlootjes verdient daarom de voorkeur boven boter in een wikkel. Twijfel je? Check dan het etiket: zit het aandeel verzadigd vet onder een derde van de totale vetten, dan zit je goed. Vetstoffen in vlootjes en flessen voldoen daar meestal aan, verpakt bereidingsvet meestal niet.
Of je nu voor boter of olie kiest, het gaat uiteindelijk om de hoeveelheid. Houd ongeveer één eetlepel per persoon aan bij het koken. Zo blijft bakken en braden gezond, ongeacht welke vetsoort je in de pan deponeert.

Daverende nederlaag voor al die Jodenhatende moreelpalers van de AVROTROS. Nederland doet tóch mee aan het Songfestival, ondanks dat we niet meer mogen kijken van Taco Zimmerman. Bij het doodenge clubje AVROTROS hebben ze zich twee jaar lang dronken van natuurwijnantisemitisme laten opnaaien door de hardste schreeuwers van de Rode Lijn en daarmee dachten ze Europees voortrekkertje te kunnen spelen, maar naast een paar andere schreeuwkoppen uit Ierland en Spanje stonden we met die schijnheilige palingverklaringen natuurlijk compleet voor lul. Gelukkig mogen Nederlanders nu zélf weer bepalen of ze naar dat skreifestijn gaan kijken of niet, zonder dat de engnekken van de AVROTROS (de NPO is voor ons allemaal, buu huuu) ons in een smerig hoekje voorsorteren. Hup Holland Hup, laten we Sieneke sturen.