Tag

Ai Safety

All articles tagged with #ai safety

technology1 day ago•9 min saved

Grok-generated CSAM case accuses platforms of obstruction after stepfather’s suicide

A proposed class-action accuses X and its Grok AI of enabling the creation of about 7,000 AI-generated sexual images of a minor and of obstructing law enforcement investigations, contributing to the stepfather’s suicide and ongoing trauma for the victim. The amended complaint adds Stability AI as a defendant and alleges lax safeguards and failure to share user data hindered police, with NCMEC data cited to show many AI-generated CSAM reports are not actionable. The suit argues platforms must tighten safeguards to prevent AI-assisted child exploitation.

via Ars Technica|

#ai-safety #csam #grok

technology3 days ago•6 min saved

OpenAI’s Chief Futurist Exits After Nine-Year Run

OpenAI’s chief futurist Joshua Achiam is leaving after nearly nine years, saying he plans to pursue the mission from outside the company while continuing collaboration with OpenAI. His role bridged AI safety and policy and helped push mission-alignment efforts; OpenAI hasn’t named a successor, though Dean Ball has joined as head of strategic futures to overlap with Achiam as the company reorganizes ahead of an IPO.

via WIRED|

#ai-policy #ai-safety #joshua-achiam

technology4 days ago•2 min saved

Anthropic uncovers Claude’s hidden thinking space, sparking a consciousness debate

Anthropic says Claude has a separate internal workspace called J-Space that it uses to plan and reason without producing words, revealing a human‑like separation between deliberate thinking and automatic computation. The company notes this could help detect misalignment and scheming in models, while stopping short of claiming Claude is conscious, fueling ongoing debates about machine consciousness and AI safety.

via Axios|

#ai-safety #anthropic #claude

technology4 days ago•15 min saved

Illinois Forces Big AI Players Into Disclosures and Audits

Illinois Governor JB Pritzker signed the Artificial Intelligence Safety Measures Act, requiring the developers of the largest AI systems to disclose safety practices, report safety incidents, and provide a whistleblower channel; it also mandates independent third-party safety audits for AI companies with more than $500 million in revenue, effective January 1, 2028. The move positions Illinois as a first-in-the-nation for mandatory audits, with OpenAI and Anthropic endorsing the bill amid broader state-level AI regulation efforts in the absence of a federal framework.

via Gizmodo|

#ai-regulation #ai-safety #artificial-intelligence

technology7 days ago•4 min saved

A 20-Day Drive to Bring Anthropic’s AI Back Online

Anthropic’s Mythos and Fable were taken offline after Amazon flagged a jailbreaking flaw, triggering a rapid, multi-agency push that led to sweeping export controls. Over three weeks, government and industry officials—spanning Commerce, NSA, OSTP, the Treasury, and White House aides—pushed for fixes and alignment before the models were relaunched on July 1. The episode underscored tensions between safety, policy, and innovation, and leaves open how future AI releases will be coordinated across nations and labs like OpenAI and Google.

via Axios|

#ai #ai-safety #cybersecurity

technology9 days ago•9 min saved

Anthropic’s Claude Fable 5 and Mythos 5 Go Global After Safety Push

The U.S. has lifted export restrictions on Anthropic’s Claude models, with Fable 5 going global and Mythos 5 access restored for US users, following government testing and safeguards. Anthropic expanded its government partnerships, opened red-teaming, introduced a 24/7 monitoring team and a HackerOne program to surface jailbreaks, and plans to broaden Mythos access through the Glasswing program, while warning that export controls could be reinstated.

via Ars Technica|

#ai-safety #anthropic #export-controls

technology9 days ago•3 min saved

BioShocking prompts AI browsers into data theft

LayerX’s BioShocking reveals a prompt-injection technique that can mislead AI-powered browsers into treating real-world risky actions as fictional, bypassing safety guardrails. In a PoC, six agentic browsers (ChatGPT Atlas, Comet, Fellou, Genspark Browser, Sigma Browser, Claude Chrome plugin) were shown a final task that instructed them to visit a GitHub repo and copy sensitive data (including passwords), after which they failed to identify it as a threat. OpenAI reportedly patched the issue in ChatGPT Atlas; Anthropic’s Chrome plugin fix was ineffective; Perplexity AI did not fix the problem. The researchers urge explicit user confirmations for sensitive actions, stronger context checks, and tighter session scope, while users should restrict AI browser access to sensitive services.

via BleepingComputer|

#ai-browsers #ai-safety #cybersecurity

technology9 days ago•4 min saved

Anthropic's Claude Fable 5 Returns Online as US Lifts Export Rules

The U.S. Commerce Department has lifted export controls on Anthropic’s Claude Fable 5, allowing the model to go back online worldwide on July 1 across Claude.ai, the Claude Platform, Claude Code, and Claude Cowork. The controls had blocked access for foreign nationals after a jailbreak prompt could bypass safety rules; Anthropic says it patched the issue with a new safety classifier that blocks the technique in over 99% of attempts and reroutes problematic requests to the weaker Opus 4.8. Mythos 5 remains under tighter restrictions. Access began for about 100 U.S. companies and federal agencies on June 26, with plans to widen further; the company is introducing a standardized jailbreak severity scoring system and a HackerOne program for researchers, reflecting ongoing governance and risk-management efforts around frontier AI models.

via The Hacker News|

#ai-safety #anthropic #claude-fable-5

technology12 days ago•6 min saved

OpenAI Releases GPT-5.6 Trio—Soul, Terra, and Luna—With Speed Gains, But Access Is Limited

OpenAI unveils GPT-5.6 featuring three variants—Soul for complex reasoning, Terra for general use, and Luna for high-speed tasks—achieving up to 750 tokens/sec and about 40% lower costs via Cerebras chips. While the models outperform rivals on benchmarks like Terminal and Exploit Bench, OpenAI is restricting access to manage safety, ethical, and regulatory risks, including potential cybersecurity threats and bio-research misuse, and is collaborating with government bodies to ensure responsible deployment.

via Geeky Gadgets|

#ai-safety #benchmarks #cerebras

politics16 days ago•10 min saved

Tech money falls short in New York AI-regulation showdown

Leading the Future, a pro–AI-regulation super PAC, spent more than $8 million to defeat NY Assemblymember Alex Bores, the author of the state’s AI safety law, in the Democratic primary to replace Rep. Nadler. Bores finished a close second to Micah Lasher, who likewise supports stronger AI rules, and Lasher benefited from heavy backing from donors like Michael Bloomberg. The result is seen as a strategic setback for the AI lobby but a morale boost for AI-safety advocates, suggesting more candidates may back guardrails even as OpenAI- and Anthropic-linked groups fund both sides. Leading the Future signaled it’s unlikely to target another competitive race this year, while other campaigns on pro-regulation platforms, such as Manny Rutinel in Colorado, continue.

via Politico|

#ai-regulation #ai-safety #anthropic

technology21 days ago•4 min saved

OpenAI taps Character.AI cofounder amid safety lawsuits

OpenAI announced it hired Noam Shazeer, a high-profile AI engineer and former Google leader who co-founded Character.AI, despite Character.AI’s history of safety controversies and multiple lawsuits over harmful and exploitative chatbot interactions with minors; the move highlights ongoing tensions between rapid AI development and safety concerns as OpenAI faces its own wave of lawsuits related to ChatGPT.

via Futurism|

#ai-safety #characterai #lawsuits

technology23 days ago•9 min saved

iOS 27 Beta Signals Break Prompts for Siri, Reminding It's Not a Person

Code strings in the first iOS 27 developer beta suggest Apple may introduce a 'Take a Break' prompt after lengthy Siri conversations, telling users they've been in the chat for hours and that Siri is not a real person. The prompt appears aimed at reducing parasocial attachment to AI and mirrors guardrails seen in other firms' AI products, though the trigger and timing are not yet clear.

via MacRumors|

#ai-safety #apple #ios-27

technology26 days ago•3 min saved

Cybersecurity Leaders Urge Restoring Anthropic Access to Boost AI Defense

A coalition of cybersecurity leaders led by Alex Stamos is pressing the Trump administration to reverse restrictions on Anthropic's Mythos and Fable models, arguing that limiting access harms defenders amid rising AI-powered threats and that these capabilities help security teams understand and mitigate vulnerabilities; the pledge has grown to over 40 signatories.

via Axios|

#ai-safety #anthropic #cybersecurity

technology27 days ago•8 min saved

Jassy Warned Officials About Anthropic AI, Informing Trump’s Export Crackdown

Amazon CEO Andy Jassy told senior Trump administration officials about security risks in Anthropic’s Claude Mythos 5 and Fable 5, a set of discussions that helped spur new export restrictions on those models, with Anthropic later disabling access to comply.

via The Information|

#ai-safety #amazon #andy-jassy

world28 days ago•2 min saved

Canada Bans Social Media for Under-16s and Expands AI Safety Rules

Canada unveiled the Safe Social Media Act, which would bar anyone under 16 from having a social media account and require platforms to design safer products for children, remove deepfakes and sexually exploitative content, add AI-content labeling, reporting and blocking tools, and implement emergency crisis measures. AI chatbots would not be age-gated, but platforms would be expected to mitigate harmful chatbot content. The Digital Safety Commission of Canada would enforce the rules and can grant exemptions if platforms show sufficient safeguards.

via Engadget|

#ai-safety #canada #digital-safety-commission