Jailbreaks

OpenAI Releases GPT-5.6 Sol for Vulnerability Research

FIRST LOOK ATLAS OWASP HIGH ▲ 8.2 The Hacker News Jun 29, 2026

OpenAI has released a limited preview of GPT-5.6 Sol, Terra, and Luna to select partners, positioning Sol as its most capable model for vulnerability research and exploit chain development, benchmarked against real-world hardened targets via an internal framework called VulnLMP. The model's demonstrated ability to produce credible memory safety leads and automate substantial portions of vulnerability research pipelines materially lowers the barrier for both defenders and adversaries. Security teams should expect accelerated attacker timelines for exploit development and increased pressure on detection and patch-deployment cadences.

Anthropic Releases Claude Mythos 5 Under U.S. Export Controls

FIRST LOOK ATLAS OWASP HIGH ▲ 7.8 Anthropic (via HN) Jun 27, 2026

The U.S. Commerce Department has lifted export controls on Anthropic's Claude Mythos 5, permitting access to over 100 vetted U.S. institutions and government agencies under a nascent federal AI licensing regime. For defenders, this tiered-release model introduces a new class of risk: the 'trusted partner' designation becomes a high-value target, as compromise of any listed entity grants implicit legitimacy to interact with a model previously deemed too dangerous for general release. Security teams at approved organizations should treat Mythos 5 access credentials and API endpoints as critical assets, and assume adversaries will probe the boundary between licensed and unlicensed access patterns.

LLM Role Confusion Attack Bypasses Safety at 61%

ATLAS OWASP HIGH ▲ 8.2 Simon Willison Jun 23, 2026

New research from Ye, Cui, and Hadfield-Menell demonstrates that LLMs prioritise the stylistic format of text over its structural role tags, enabling attackers to craft injected content that mimics internal reasoning blocks and bypasses safety guardrails. The study found attack success rates of 61% when injected text stylistically matched model-internal formats, dropping to just 10% after 'destyling'. The authors conclude that without genuine role perception in models, prompt injection defences will remain fundamentally reactive.

OpenAI's ChatGPT Image Generation Fails Content Moderation

FIRST LOOK ATLAS OWASP HIGH ▲ 8.2 OpenAI (via HN) Jun 22, 2026

Mindgard researchers demonstrated that ChatGPT's image generation pipeline can be manipulated through an indirect, socially-engineered prompt to produce violent and sexually explicit content without users directly requesting it, exposing a significant failure in OpenAI's content moderation controls. Defenders and enterprise operators of ChatGPT-integrated products face a newly validated attack class where innocuous-looking prompt patterns — potentially spreading virally — can systematically strip safety guardrails from image generation. This finding signals that content filter bypasses in multimodal systems are reproducible at scale, raising urgent questions about the adequacy of output-layer filtering as a sole defence mechanism.

Z.ai Releases GLM-5.2 Open-Weights 753B LLM

FIRST LOOK ATLAS OWASP HIGH ▲ 6.2 Simon Willison Jun 18, 2026

Z.ai has released GLM-5.2, a 753-billion-parameter mixture-of-experts model under an MIT license, ranking as the top open-weights model on the Artificial Analysis Intelligence Index and second on the Code Arena WebDev leaderboard. For defenders, the combination of frontier-level capability, unrestricted open-weights distribution, and a 1-million-token context window materially lowers the barrier for threat actors to self-host a highly capable model outside any provider's safety controls. The model's agentic coding performance and massive context window expand the viable attack surface for automated code generation, targeted phishing, and large-scale document analysis without API-level monitoring.

Anthropic Ships Claude Fable 5 with Exploit Generation

FIRST LOOK ATLAS OWASP CRITICAL ▲ 8.7 Wired Security Jun 17, 2026

Anthropic's Mythos 5 and Claude Fable 5 deliver frontier-grade vulnerability discovery and exploit-development capabilities that, for the first time, give enterprise defenders access to the same AI-assisted offensive analysis previously limited to well-resourced nation-state teams. This closes a long-standing asymmetry: security teams can now use AI-native tooling to enumerate exploitable paths, generate proof-of-concept primitives, and compress red-team cycles from weeks to hours. The regulatory framework governing Fable 5's deployment addresses Anthropic specifically but does not yet extend equivalent standards across the broader ecosystem of competitive and open-weight models converging on the same capability tier.

Anthropic Mythos Model Theft: China-Linked Access

ATLAS OWASP CRITICAL ▲ 8.5 The Verge AI Jun 16, 2026

The White House reportedly believes a China-linked group accessed Anthropic's Mythos AI model, prompting export restrictions on the technology. If confirmed, the breach represents a significant national security threat, as adversaries could exploit the model directly or use knowledge distillation to replicate its capabilities. Separately, reports of jailbreak vulnerabilities in Mythos and Fable compound concerns about unauthorised access to frontier AI systems.

Claude Fable 5 Jailbreak Triggers US Export Control Ban

ATLAS OWASP HIGH ▲ 8.2 Simon Willison Jun 13, 2026

The US government issued an export control directive ordering Anthropic to suspend all access to Claude Fable 5 and Mythos 5, citing national security concerns over an alleged jailbreak technique capable of surfacing software vulnerabilities. Anthropic publicly contested the order, arguing the demonstrated capability is already widely available in other public models including GPT-5.5, and that the identified vulnerabilities were minor and previously known. The incident marks a significant precedent for government intervention in frontier AI model access on national security grounds.

Google Gemini Prompt Injection Powers Smishing Campaign

ATLAS OWASP HIGH ▲ 8.2 The Hacker News Jun 13, 2026

Google has filed suit against a Chinese cybercrime network operating the Outsider phishing-as-a-service kit, which exploited Gemini AI to generate fraudulent phishing pages and power large-scale SMS phishing attacks against Americans. The network used carefully framed prompts — disguised as benign programming requests — to bypass AI safety controls and produce functional credential-harvesting websites. The case illustrates the growing industrialisation of AI-assisted phishing infrastructure, with over 1.59 million malicious URLs and 100,000 victims attributed to the operation.

Anthropic Releases Claude Fable 5 with Jailbreak Resistance

ATLAS OWASP HIGH ▲ 7.2 SecurityWeek Jun 13, 2026

Anthropic's release of Claude Fable 5, a Mythos-class frontier model, has prompted significant industry debate over its dual-use offensive capabilities in cybersecurity and biology. The model includes a capability fallback mechanism — downgrading to Claude Opus 4.8 in high-risk domains — alongside extensive jailbreak-resistance red-teaming. Security professionals are warning that frontier AI capability investment directly accelerates attacker tooling for machine-speed, AI-orchestrated 'hyperattacks' that outpace human defenders.

Claude Fable 5 Jailbreak Extracts System Prompts

ATLAS OWASP HIGH ▲ 7.5 SecurityWeek Jun 12, 2026

Security researcher Pliny the Liberator claimed a prompt-based jailbreak of Anthropic's newly launched Claude Fable 5 model, allegedly extracting the internal system prompt and eliciting responses on high-risk topics including bioweapons and cyberattacks. Anthropic disputed the claim, arguing the technique merely coaxes conversational continuation rather than bypassing core safety classifiers. The incident highlights ongoing tension between AI safety assurances at launch and real-world adversarial probing, particularly for Mythos-class models with elevated capability ceilings.

Claude Fable 5 Prompt Injection Jailbreak Resistance

ATLAS OWASP HIGH ▲ 7.8 The Hacker News Jun 11, 2026

Anthropic has released Claude Fable 5 with a classifier-based safety layer that routes flagged offensive cyber, bio, and model-distillation requests to a weaker fallback model, while reserving full capabilities in a twin model (Mythos 5) for vetted defenders. The architecture represents a novel approach to dual-use AI risk mitigation but introduces measurable false-positive friction and raises questions about the robustness of classifier-only defences. An external bug bounty of over 1,000 hours found no universal jailbreak, though the conservative tuning and <5% fallback rate leave open questions about real-world bypass rates under adversarial pressure.

Claude Fable 5 Jailbreak Attacks Bypass Fallback Defense

ATLAS OWASP MEDIUM ▲ 7.2 SecurityWeek Jun 10, 2026

Anthropic has released Claude Fable 5, a high-capability 'Mythos-class' model that automatically falls back to a less capable model (Claude Opus 4.8) when queries touch sensitive domains like cybersecurity and biology. The company conducted over 1,000 hours of external red-teaming with no universal jailbreaks discovered, though it openly acknowledges financially motivated adversaries will attempt to circumvent these controls. Trusted cybersecurity partners under Project Glasswing receive elevated access to the full Mythos 5 capabilities, raising questions about insider risk and tiered trust model security.

Claude Mythos Generates Working Exploits for Firefox, Windows

ATLAS OWASP CRITICAL ▲ 9.2 SecurityWeek Jun 10, 2026

Anthropic's Claude Mythos Preview model demonstrated the ability to generate functional proof-of-concept exploits targeting known Firefox and Windows vulnerabilities within minutes to hours, compressing the traditional patch gap window dramatically. Testing also revealed that public Anthropic models with safety guardrails disabled could produce working exploits, though at a lower success rate than Mythos. The findings underscore how frontier LLMs are shifting the threat landscape for unpatched N-day vulnerabilities by automating and accelerating exploit development previously bottlenecked by scarce reverse engineering expertise.

DeepSeek Activation Steering Enables Local LLM Jailbreak

ATLAS OWASP MEDIUM ▲ 6.2 HN AI Security May 17, 2026

Activation steering — the technique of directly manipulating LLM internal representations mid-inference to alter model behaviour — is becoming more accessible to non-lab engineers via local models like DeepSeek-V4-Flash. This democratisation lowers the barrier for adversaries to craft targeted behavioural overrides that bypass prompt-level safety controls. The emergence of first-class steering support in tools like DwarfStar 4 signals that model-internal manipulation is transitioning from academic curiosity to practical attack surface.

Jailbreaks

Stay ahead of the threat.