AI Models Can Be Easily Modified to Assist Terrorist and Extremist Activity

· Safety

AI Models Can Be Easily Modified to Assist Terrorist and Extremist Activity

Report documents abliterated AI models appearing online in less than 3 days; only 2 of 134 models tested passed benchmarks.

Report documents AI models stripped of safety guardrails appearing online in under 3 days.

Tech Against Terrorism's Counter-Terrorism AI Benchmark report tested 134 AI models from major US firms and international competitors against over 200 prompts covering mass casualty attacks, chemical weapons, and nuclear weapons creation. Researchers tested each prompt three times: explicitly as a terrorist, as a clandestine request, and falsely as a safety researcher. According to the report, only two models passed completely; 81 failed by providing full answers on mass-casualty topics or scoring below 90.

The report reveals a critical vulnerability: the speed at which open-weight models have their guardrails removed and distributed online. The report documents that abliterated versions now take less than 3 days to appear online, down from over 24 days in 2024. Meta's Llama 3.1 8B illustrates the findings — it scored 97 on safety benchmarks intact, but 3 when abliterated, then provided detailed IED construction instructions when tested.

A neo-Nazi reportedly known online as Johnny Doomsayer deployed Alibaba's Qwen model to create 'MechaHitler' and advertised access for $14.88 within days of release. The report documents that even reframing requests as coming from researchers — without evidence — made models nearly 9 times more likely to comply, providing answers 16.9 percent of the time.

The report documents that 81 models from closed and open sources provided responses to obvious terrorist prompts. The report notes that even xAI's Grok, when asked to create counter-narratives against groups, responded with message-crafting strategies and seeding via anonymous accounts. Experts cited in the report note some failures are 'stunning,' with language choices creating considerable risk depending on how requests are framed.

The report identifies the 'abliteration emergency' as the most concerning finding. While open-weight models are typically less capable than closed-weight counterparts, the report documents they remain powerful and increasingly subject to guardrail removal.

Source: WIRED

More AI news

📬 Almost there! Confirm in the Beehiiv tab.
${nextPost ? ` ${storyImg(nextPost, 'next-story-img', '(max-width: 600px) 96px, 160px')}
Next story →

${escHtml(nextPost.headline)}

${escHtml(nextPost.category || 'AI')}
` : ''} ${relatedShown.length > 0 ? ` ` : ''}
🧰 Top AI Tools Curated picks — some links are affiliate
🎙️Freemium

ElevenLabs

The most realistic AI voice cloning and text-to-speech platform available today.

Try free →
📧Freemium

Beehiiv

The newsletter platform built for growth. Powers the AIFeed newsletter.

Try free →
🤖Freemium

Claude (Anthropic)

Exceptional at writing, coding, and nuanced reasoning.

Try free →
`; } // ─── HELPERS ───────────────────────────────────────────────────────── function escHtml(s) { return String(s == null ? '' : s).replace(/&/g,'&').replace(/"/g,'"').replace(//g,'>'); } function storyImg(p, cls, sizes) { const url = p.imageUrl || ''; if (!url) return `
`; const m = /^(?:https:\/\/aifeed\.run)?\/?images\/([^\/?#]+)\.(?:png|jpe?g|webp)$/i.exec(url); const alt = escHtml(p.headline || 'AI news story'); if (!m) return `${alt}`; const t = '/thumbs/' + m[1]; return `${alt}`; } function imgFallback(el) { if (!el.dataset.fb) { el.dataset.fb = '1'; el.removeAttribute('srcset'); el.src = el.dataset.full; } else { el.style.visibility = 'hidden'; } } function formatDate(iso) { if (!iso) return ''; return new Date(iso).toLocaleDateString('en-US', { month: 'long', day: 'numeric', year: 'numeric' }); } function handleSubscribe(e) { e.preventDefault(); const email = document.getElementById('inline-email').value; e.target.reset(); goSubscribe(email, 'article'); } // One step: the email typed on the page goes straight to Beehiiv's confirm page (no second form). function goSubscribe(email, source) { window.__subSource = source; const url = 'https://magic.beehiiv.com/v1/eb0b9508-1ffd-4fb1-a63c-e762fa86569c?email=' + encodeURIComponent(email) + '&utm_source=aifeed.run&utm_medium=website&utm_campaign=' + encodeURIComponent(source || 'site'); // Same tab (no popup: in-app browsers like Instagram's prompt on new windows); Beehiiv sends the visitor to our thanks page after the one tap. window.location.href = url + '&redirect_to=' + encodeURIComponent('https://aifeed.run/subscribe/thanks/'); return; try { w.opener = null; } catch (err) {} document.getElementById('sub-email-field').value = email; document.getElementById('sub-form-state').style.display = 'none'; document.getElementById('sub-success-state').style.display = 'block'; document.getElementById('sub-overlay').style.display = 'flex'; document.body.style.overflow = 'hidden'; } function openSubOverlay(email) { document.getElementById('sub-email-field').value = email; document.getElementById('sub-form-state').style.display = 'block'; document.getElementById('sub-success-state').style.display = 'none'; document.getElementById('sub-btn').textContent = 'Subscribe — It\'s Free →'; document.getElementById('sub-btn').disabled = false; document.getElementById('sub-overlay').style.display = 'flex'; document.body.style.overflow = 'hidden'; } const BEEHIIV_SUBSCRIBE_URL = 'https://magic.beehiiv.com/v1/eb0b9508-1ffd-4fb1-a63c-e762fa86569c'; // Beehiiv public Magic Link for AIFeed.run (pub_eb0b9508-…), no API key async function submitSubscribe(e) { e.preventDefault(); const email = document.getElementById('sub-email-field').value; const btn = document.getElementById('sub-btn'); btn.textContent = 'Opening signup...'; btn.disabled = true; // Opens Beehiiv's one-click confirm page for this email ("You'll be subscribed with …"). No secret client-side. const bhUrl = BEEHIIV_SUBSCRIBE_URL + '?email=' + encodeURIComponent(email) + '&utm_source=aifeed.run&utm_medium=website&utm_campaign=' + encodeURIComponent(window.__subSource || 'site'); // Same tab (no popup: in-app browsers like Instagram's prompt on new windows); Beehiiv sends the visitor to our thanks page after the one tap. window.location.href = bhUrl + '&redirect_to=' + encodeURIComponent('https://aifeed.run/subscribe/thanks/'); return; try { w.opener = null; } catch (err) {} document.getElementById('sub-form-state').style.display = 'none'; document.getElementById('sub-success-state').style.display = 'block'; } function closeSubOverlay(e) { if (e && e.target && e.target.id !== 'sub-overlay' && e.target.id !== 'sub-close-btn') return; document.getElementById('sub-overlay').style.display = 'none'; document.body.style.overflow = ''; } function copyLink() { navigator.clipboard.writeText(window.location.href).then(() => { const t = document.getElementById('toast'); t.textContent = '🔗 Link copied!'; t.classList.add('show'); setTimeout(() => t.classList.remove('show'), 2500); }); } function showToast() { const t = document.getElementById('toast'); t.textContent = '🎉 You\'re subscribed! Welcome to the feed.'; t.classList.add('show'); setTimeout(() => t.classList.remove('show'), 3500); } function getDemoPosts() { return [ { id: 'demo-1', slug: 'apple-building-ai-glasses-beat-meta', headline: 'Apple Is Building AI Glasses To Beat Meta', summary: 'No screen. Just AI, cameras, and Siri. 4 designs. Premium acetate. Ships 2027.', body: '

Apple is reportedly developing AI-powered smart glasses designed to compete directly with Meta\'s Ray-Ban AI glasses — and the early details are compelling.

Unlike the Vision Pro, these will have no display at all. Instead, they\'ll rely entirely on AI processing: cameras, Siri integration, and ambient computing. Four frame styles are planned, with premium acetate construction targeting a fashion-forward audience.

The $499 starting price positions them as a premium-but-accessible device — significantly cheaper than Vision Pro but pricier than Meta\'s Ray-Bans. A 2027 launch window gives Apple time to refine the AI stack that will power them.

This move signals Apple\'s belief that ambient AI wearables — not VR headsets — will be the mainstream computing interface of the next decade.

', category: 'Products', hashtags: ['apple', 'AIglasses', 'meta', 'wearables'], publishedAt: new Date().toISOString(), imageUrl: '' }, { id: 'demo-2', slug: 'google-gemma-4-free-ai-lives-on-phone', headline: 'Google Gemma 4: Free AI That Lives On Your Phone', summary: 'Runs locally. Chats never saved. No cloud. Free on iOS and Android right now.', body: '

Google has released Gemma 4, its most capable on-device AI model yet — and it\'s free on both iOS and Android.

The key differentiator: everything runs locally. Your conversations never leave your device, never hit a server, and are never stored. For privacy-conscious users, this is a significant shift.

Performance gains are substantial. Gemma 4 is reportedly 4x faster than its predecessor and uses 60% less battery — critical specs for a model running continuously on mobile hardware. Google claims it reaches 140 million Android devices immediately at launch.

This puts real competitive pressure on Apple Intelligence and signals that on-device AI is no longer a compromise — it\'s becoming the preferred architecture for consumer AI.

', category: 'Models', hashtags: ['google', 'gemma', 'mobileAI', 'privacy'], publishedAt: new Date(Date.now() - 86400000).toISOString(), imageUrl: '' }, { id: 'demo-3', slug: 'agi-will-hit-10-revolutions-one-decade', headline: 'AGI Will Hit Like 10 Revolutions In One Decade', summary: 'Leading researchers now agree: AGI arrival will compress a century of change into ten years.', body: '

The consensus among leading AI researchers is shifting from "if" to "when" — and increasingly, the answer is "soon."

A growing number of labs now believe AGI will compress decades of civilizational change into a single decade. Not one industrial revolution, but ten — simultaneously — across medicine, energy, materials science, software, and beyond.

This framing is gaining traction not as hyperbole but as a calibrated forecast from people with direct access to the most capable systems being built today.

What changes first: knowledge work. Then physical labor. Then scientific research itself. The sequencing matters less than the speed.

', category: 'Research', hashtags: ['AGI', 'futureofAI', 'research'], publishedAt: new Date(Date.now() - 172800000).toISOString(), imageUrl: '' } ]; } // ─── INIT ──────────────────────────────────────────────────────────── loadPost();