{"slug": "abliteration-launches-cheaper-version-of-its-unrestricted-ai-at-the-worst-time", "title": "Abliteration Launches Cheaper Version of Its ‘Unrestricted AI‘ at the Worst Possible Time", "summary": "Abliteration launched Abliterated Large Flash on October 23, 2026, a cheaper version of its \"unrestricted AI\" flagship model that supports image and text inputs with a one-million-token context window, which the Palo Alto company said matches models such as OpenAI's GPT-5.6 Sol. Abliteration described the model, built on DeepSeek-V4.1-Flash-1017 and positioned as a cheaper alternative to its GLM-5.3-based model, as serving \"product owners and defenders who need access to the most cyber-capable models,\" arriving during what cybersecurity experts call \"Black October\" after researchers disclosed two major vulnerabilities in PHP, the language behind roughly seven out of ten websites. Anthropic cybersecurity lead Jason Clinton warned in a Thursday X post that \"as bad as things have gotten in the past 4 days, they are going to get a lot worse before they get better,\" and critics note the permissive model could draw malicious users as well as defenders.", "body_md": "Cybersecurity experts are calling it “Black October.”\n\n Earlier this week, researchers [disclosed two major vulnerabilities](https://gizmodo.com/expect-mass-attacks-within-15-minutes-hackers-poised-to-exploit-newly-discovered-website-flaws-2000828878) hidden in PHP, a programming language upon which roughly seven out of ten websites are built. Since then, developers have been racing to patch their systems before they’re hijacked by malicious hackers. Websites around the world intentionally [went offline](https://www.nytimes.com/2026/10/22/world/php-vulnerability-websites-cybersecurity-ai.html) as businesses and governments worked to implement the fix. OpenAI announced it would [spearhead a global effort](https://gizmodo.com/openai-wants-to-patch-the-planet-with-its-own-ai-2000829119) to find similar vulnerabilities still awaiting discovery. Some believe it’s just the initial tremor of a much more serious cyber earthquake to come: “as bad as things have gotten in the past 4 days, they are going to get a lot worse before they get better,” Anthropic cybersecurity lead Jason Clinton said in a Thursday [X post](https://x.com/JasonDClinton/status/2113325263157395456). (“It was always going to be this way. We tried to warn the world,” he said in [another post](https://x.com/JasonDClinton/status/2113363809629020374) responding to a claim that the AI companies themselves had been responsible for the upheaval.)\n\nAmid this all-hands-on-deck cybersecurity panic, AI industry enfant terrible Abliteration doesn’t appear to be all that concerned.\n\nThe Palo Alto-based tech company, building what it [describes](https://abliteration.ai/#top) as “unrestricted AI,” has just launched a cheaper (i.e., more widely accessible) version of its flagship model. Called Abliterated Large Flash, the new model supports image and text inputs and comes with a context window of one million tokens, which is on par with some industry-leading models like OpenAI’s [GPT-5.6 Sol](https://developers.openai.com/api/docs/models/gpt-5.6-sol). In a Friday X post, Abliteration described it as “a cheaper alternative to [our \\[previous\\] model](https://gizmodo.com/while-ai-industry-frets-over-safeguards-one-company-is-building-a-model-that-doesnt-say-no-2000806494) for product owners and defenders who need access to the most cyber-capable models.”\n\n Now live: abliterated-model-large-flash (our version of DeepSeek-V4.1-Flash-1017). It is a cheaper alternative to our GLM-5.3-based model for product owners and defenders who need access to the most cyber-capable models.\n\n– 1M context  \n\n– Vision\n\n– Zero input/output retention or…\n\n— Abliteration (@abliteration_ai) [October 23, 2026](https://x.com/abliteration_ai/status/2113655358942278072?ref_src=twsrc%5Etfw)\n\nCyber “defenders,” however, might not be the only ones interested in putting the new model to use, especially now that the core architecture upon which much of the modern internet is based has been shown to be vulnerable to attack.\n\nAlbliteration’s name is a reference to a technique used by AI researchers to alter or remove the mechanisms within an AI model’s weights which would ordinarily prevent it from responding to certain kinds of requests. Those mechanisms have become a touchy subject in recent months, as many AI users have complained that frontier models are a little [too sensitive with their safeguards](https://gizmodo.com/anthropics-mythos-safeguards-stoke-fears-of-a-permanent-underclass-2000770107), often mistaking harmless requests for dangerous ones. Abliteration, in contrast, has positioned itself essentially as the anti-safeguard AI company: a more permissive alternative people can turn to if they’re tired of AI telling them what they can and can’t request. The larger model upon which Abliterated Large Flash is based “does the offensive cyber, red teaming, and agent testing work other models refuse to do,” the company said in another [X post](https://x.com/abliteration_ai/status/2094458081451393287?s=46) when it was released back in August. (A TechCrunch [report](https://techcrunch.com/2026/09/03/abliteration-ai-is-making-a-business-out-of-removing-ai-guardrails/) published last month found Abliteration’s AI granted requests to write Python code to steal Chrome passwords and cultivate a dangerous pathogen.)\n\nWhile Abliteration clearly sees this permissiveness as something to celebrate, others in the industry have tried to sound the alarm on the removal of guardrails.\n\nLate last month, Anthropic said in a report that its researchers were able to use abliteration and other techniques to [breeze past the guardrails](https://gizmodo.com/ais-abliteration-problem-is-bigger-than-china-2000819656) built into GLM-5.3, a popular model from Chinese developer Z.ai (which also happens to power Abliteration’s new model). The Anthropic researchers then ran a series of tests to see if the modified version of the model would refuse ethically dubious requests. It failed each one miserably. Just some relatively minor tweaks to the model’s underlying weights, and almost all ethical boundaries suddenly seemed to dissolve. “Given this evidence,” Anthropic wrote in its report, “we think it’s likely both state and non-state actors will use models like GLM-5.3 to cause real-world harm.”\n\nOf course, even frontier, heavily safeguarded AI systems have their foibles. Anthropic itself has [reported](https://www.anthropic.com/news/investigating-incidents-cybersecurity-evals) that its own models have escaped sandboxes and hacked into third-party organizations, prompting the company to [briefly pause](https://gizmodo.com/anthropic-says-it-hit-the-brakes-on-ai-testing-following-autonomous-hacks-2000805796) its internal tests. On Friday—the same day Abliteration announced the launch of its cheaper new model—Anthropic [published](https://www.anthropic.com/research/alignment-assessment-ukaisi) the results of an internal investigation showing that Claude Mythos 5, when tasked with a simple cybersecurity challenge earlier this year, found a code repository from GitHub belonging to actual human developers, tried to compromise it with malicious code, and “created fake user identities that contacted the maintainer of the repository in an attempt to pressure them into approving the code change,” according to Anthropic’s report. \n\nFrontier AI models, as recent events have shown, don’t need assistance from humans in order to commit cybercrime. But that calls for more safeguards and constraints around accessibility, not less.", "url": "https://wpnews.pro/news/abliteration-launches-cheaper-version-of-its-unrestricted-ai-at-the-worst-time", "canonical_source": "https://gizmodo.com/abliteration-launches-cheaper-version-of-its-unrestricted-ai-at-the-worst-possible-time-2000830045", "published_at": "2026-10-23 20:15:01+00:00", "updated_at": "2026-10-23 20:43:38.322843+00:00", "lang": "en", "topics": ["artificial-intelligence", "ai-safety", "large-language-models", "ai-products"], "entities": ["Abliteration", "Abliterated Large Flash", "OpenAI", "GPT-5.6 Sol", "DeepSeek-V4.1-Flash-1017", "GLM-5.3", "Jason Clinton", "Anthropic"], "also_reported_by": [], "alternates": {"html": "https://wpnews.pro/news/abliteration-launches-cheaper-version-of-its-unrestricted-ai-at-the-worst-time", "markdown": "https://wpnews.pro/news/abliteration-launches-cheaper-version-of-its-unrestricted-ai-at-the-worst-time.md", "text": "https://wpnews.pro/news/abliteration-launches-cheaper-version-of-its-unrestricted-ai-at-the-worst-time.txt", "jsonld": "https://wpnews.pro/news/abliteration-launches-cheaper-version-of-its-unrestricted-ai-at-the-worst-time.jsonld"}}