Summarize and humanize this content to 2000 words in 6 paragraphs in English
Safety concerns about AI are growing, and now, a common method for catching misbehaving systems could have flaws, new research suggests.
In experiments with AI agents, researchers found that chain-of-thought monitoring, in which one AI checks another’s work, became far less reliable when the monitored AI’s reasoning was the main clue that something was wrong, they report August 1 on arXiv.org. That exposes a weakness in the approach: If suspicious behavior is visible mainly in the reasoning, an innocent-looking chain of thought can make that behavior much harder to catch.
Sign up for our newsletter
We summarize the week’s scientific breakthroughs every Thursday. input[type=”radio”].svelte-wecxzz{display:none}.payment-option-wrapper.svelte-wecxzz.svelte-wecxzz{display:flex;flex:1 0 28%;width:100%}.payment-option-wrapper.svelte-wecxzz.svelte-wecxzz:only-child{max-width:325px}.box-container.svelte-wecxzz.svelte-wecxzz{display:flex;flex-direction:column;flex:1 1 0;width:0;text-align:center}.inner-box.svelte-wecxzz.svelte-wecxzz{border:var(–zf-payment-options-border);border-radius:var(–zf-payment-options-borderRadius);background-color:var(–zf-payment-options-bgColor);height:100%;display:flex;flex-direction:column}.box-container.highlighted.svelte-wecxzz .inner-box.svelte-wecxzz{border-top-left-radius:0;border-top-right-radius:0}.box-container.svelte-wecxzz .box-content.svelte-wecxzz{flex-grow:1;padding:30px 20px;margin:0 10px;overflow:hidden}.box-container.svelte-wecxzz .styled-box-content.svelte-wecxzz{color:var(–zephr-color-text-main);font-family:var(–zf-root-fontFamily)}
.box-container .styled-box-content h1,
.box-container .styled-box-content h2,
.box-container .styled-box-content h3,
.box-container .styled-box-content p,
.box-container .styled-box-content span
{color:var(–zephr-color-text-main);font-family:var(–zf-root-fontFamily)}.payment-options-container .box-container .styled-box-content button{color:#fff;background-color:var(–zephr-color-action-main);border-width:calc(var(–zephr-button-borderWidth) * 1px);border-color:var(–zephr-color-action-tinted);border-radius:calc(var(–zephr-button-borderRadius) * 1px);font-size:var(–zf-button-fontSize)}.payment-options-container .box-container .styled-box-content button:hover{background-color:var(–zephr-color-action-tinted);border-color:var(–zephr-color-action-tinted)}.payment-options-container .box-container .box-content img{max-width:210px;height:auto}.box-top.svelte-wecxzz.svelte-wecxzz{height:40px}.box-container.highlighted.svelte-wecxzz .box-top.svelte-wecxzz{color:var(–zf-payment-options-highlighted-color);background-color:var(–zf-payment-options-highlighted-bgColor);position:relative;border:var(–zf-payment-options-border);border-bottom-width:0;border-radius:var(–zf-payment-options-borderRadius);border-bottom-left-radius:0;border-bottom-right-radius:0;display:flex;justify-content:center;align-items:center}.box-container.highlighted.svelte-wecxzz .box-top .box-top-text.svelte-wecxzz{font-weight:bold}.box-container.svelte-wecxzz .box-bottom.svelte-wecxzz{min-height:60px;display:flex;flex-direction:column;align-items:center;padding:20px 0 30px;margin:0 30px}.button-sub-text.svelte-wecxzz.svelte-wecxzz{margin-top:10px}
.payment-options-container.svelte-15oxdku{flex-wrap:wrap;font-size:var(–zf-payment-options-fontSize);width:100%;color:var(–zf-payment-options-color);line-height:18px;padding:31px 0 20px;display:grid;justify-content:space-evenly;justify-items:center;grid-template-columns:repeat(auto-fit, minmax(300px, 1fr));gap:13px}@media(max-width: 340px){.payment-options-container.svelte-15oxdku{grid-template-columns:repeat(auto-fit, minmax(280px, 1fr))}}
.zephr-form-input-inner-container.svelte-127vcsp{position:relative}.zephr-form-input-full-width.svelte-127vcsp{width:100%}.zephr-form-input.svelte-127vcsp{width:100%;display:block;height:calc(var(–zephr-input-height) * 1px);padding-left:8px;font-size:var(–zf-input-fontSize);border:calc(var(–zephr-input-borderWidth) * 1px) solid
var(–zephr-input-borderColor);border-radius:calc(var(–zephr-input-borderRadius) * 1px);transition:border-color 0.25s ease, box-shadow 0.25s ease;outline:0;color:var(–zephr-color-text-main);background-color:var(–zf-input-bgColor);font-family:var(–zf-root-fontFamily)}.zephr-form-input.svelte-127vcsp::placeholder{color:var(–zephr-color-background-tinted);font-family:var(–zf-root-fontFamily)}.zephr-form-input-checkbox.svelte-127vcsp{width:auto;height:auto;margin:8px 5px 0 0;float:left}.zephr-form-input-radio.svelte-127vcsp{position:absolute;opacity:0;cursor:pointer;height:0;width:0}.zephr-form-input-color[type=”color”].svelte-127vcsp{width:50px;padding:0;border-radius:50%}.zephr-form-input-color[type=”color”].svelte-127vcsp::-webkit-color-swatch{border:none;border-radius:50%;padding:0}.zephr-form-input-color[type=”color”].svelte-127vcsp::-webkit-color-swatch-wrapper{border:none;border-radius:50%;padding:0}.zephr-form-input.disabled.svelte-127vcsp,.zephr-form-input.disabled.svelte-127vcsp:hover{border:calc(var(–zephr-input-borderWidth) * 1px) solid
var(–zephr-input-borderColor);background-color:var(–zephr-color-background-tinted)}.zephr-form-input.error.svelte-127vcsp{border:1px solid var(–zephr-color-warning-main)}
.zephr-form-input-label.svelte-vnzoa4.svelte-vnzoa4{margin-top:10px;display:block;line-height:30px;font-size:var(–zf-subtext-fontSize);color:var(–zephr-color-text-tinted);font-family:var(–zf-root-fontFamily)}.zephr-form-input-label.svelte-vnzoa4.svelte-vnzoa4:first-child{margin-top:0}.zephr-form-input-label.svelte-vnzoa4 span.svelte-vnzoa4{display:block}
.zephr-form-password-match.svelte-11tz5c7.svelte-11tz5c7{margin:20px 0;justify-content:center}.zephr-form-password-match.svelte-11tz5c7 ul.svelte-11tz5c7{padding-left:0;margin:0}.zephr-form-password-requirement.svelte-11tz5c7.svelte-11tz5c7{display:flex;align-items:center;color:var(–zephr-color-text-tinted);font-size:var(–zf-subtext-fontSize);min-height:20px;font-family:var(–zf-root-fontFamily)}.zephr-form-password-requirement-icon.svelte-11tz5c7.svelte-11tz5c7{margin-right:10px;font-size:15px}
.zephr-form-password-progress.svelte-1b6w0g8.svelte-1b6w0g8{display:flex;margin-top:10px}.zephr-form-password-bar.svelte-1b6w0g8.svelte-1b6w0g8{width:100%;height:4px;border-radius:2px}.zephr-form-password-bar.svelte-1b6w0g8.svelte-1b6w0g8:not(:first-child){margin-left:8px}.zephr-form-password-requirements.svelte-1b6w0g8.svelte-1b6w0g8{margin:20px 0;justify-content:center}.zephr-form-password-requirements.svelte-1b6w0g8 ul.svelte-1b6w0g8{padding-left:0;margin:0}.zephr-form-password-requirement.svelte-1b6w0g8.svelte-1b6w0g8{display:flex;align-items:center;color:var(–zephr-color-text-tinted);font-size:var(–zf-subtext-fontSize);min-height:20px;font-family:var(–zf-root-fontFamily)}.zephr-form-password-requirement-icon.svelte-1b6w0g8.svelte-1b6w0g8{margin-right:10px;font-size:15px}
#payment-element.svelte-1oygrrs{margin-top:1em;margin-bottom:1em;border:calc(var(–zephr-input-borderWidth) * 1px) solid
var(–zephr-input-borderColor);border-radius:calc(var(–zephr-input-borderRadius) * 1px);padding:1em}
.check-image.svelte-12ljnq4{height:1.2em;vertical-align:middle;margin-right:0.5em}.zephr-form-success-message.svelte-12ljnq4{color:var(–zephr-color-text-main);display:inline-block}
.zephr-form-verification-resend-link-container.svelte-b6r3ph{margin-bottom:2em}.zephr-form-verification-resend-link.svelte-b6r3ph,.zephr-form-verification-resend-text.svelte-b6r3ph{float:right;position:absolute;padding-right:2px;padding-top:0.5em;padding-left:0.5em;font-size:var(–zf-subtext-fontSize);right:0}.zephr-form-verification-resend-link.svelte-b6r3ph{cursor:pointer;color:var(–zf-link-color)}
.zephr-form-custom-text.svelte-12fdogm{margin-top:1em;margin-bottom:1em;color:var(–zephr-color-text-main)}
.zephr-form-checkbox.svelte-eyzg32.svelte-eyzg32.svelte-eyzg32{position:absolute;opacity:0;cursor:pointer;height:0;width:0}.zephr-form-checkbox-label.svelte-eyzg32.svelte-eyzg32.svelte-eyzg32{display:flex;align-items:center;font-family:var(–zf-root-fontFamily)}.zephr-form-checkmark.svelte-eyzg32.svelte-eyzg32.svelte-eyzg32{position:relative;box-sizing:border-box;height:23px;width:23px;background-color:#fff;border:calc(var(–zephr-input-borderWidth) * 1px) solid
var(–zephr-input-borderColor);border-radius:6px;margin-right:12px;cursor:pointer}.zephr-form-checkmark.checked.svelte-eyzg32.svelte-eyzg32.svelte-eyzg32{border-color:var(–zephr-color-action-main)}.zephr-form-checkmark.checked.svelte-eyzg32.svelte-eyzg32.svelte-eyzg32:after{content:””;position:absolute;width:6px;height:13px;border:solid var(–zephr-color-action-main);border-width:0 2px 2px 0;transform:rotate(45deg);top:3px;left:8px;box-sizing:border-box}.zephr-form-checkmark.disabled.svelte-eyzg32.svelte-eyzg32.svelte-eyzg32{border:calc(var(–zephr-input-borderWidth) * 1px) solid
var(–zephr-color-background-tinted)}.zephr-form-checkmark.disabled.checked.svelte-eyzg32.svelte-eyzg32.svelte-eyzg32:after{border:solid var(–zephr-color-background-tinted);border-width:0 2px 2px 0}.zephr-form-checkmark.error.svelte-eyzg32.svelte-eyzg32.svelte-eyzg32{border:calc(var(–zephr-input-borderWidth) * 1px) solid
var(–zephr-color-warning-main)}.zephr-form-checkbox.svelte-eyzg32:focus-visible+.zephr-form-checkbox-label.svelte-eyzg32>.zephr-form-checkmark.svelte-eyzg32{outline:2px solid var(–zephr-color-action-main);outline-color:-webkit-focus-ring-color}
.zephr-form-input-radio.svelte-17rnpjv.svelte-17rnpjv.svelte-17rnpjv{position:absolute;opacity:0;cursor:pointer;height:0;width:0}.zephr-form-radio-label.svelte-17rnpjv.svelte-17rnpjv.svelte-17rnpjv{display:flex;align-items:center;font-family:var(–zf-root-fontFamily)}.zephr-form-radio-dot.svelte-17rnpjv.svelte-17rnpjv.svelte-17rnpjv{display:block;position:relative;box-sizing:border-box;height:23px;width:23px;background-color:#fff;border:1px solid #ebebeb;border-radius:50%;margin-right:12px;cursor:pointer}.checked.svelte-17rnpjv.svelte-17rnpjv.svelte-17rnpjv{border-color:var(–zephr-color-action-main)}.checked.svelte-17rnpjv.svelte-17rnpjv.svelte-17rnpjv:after{content:””;position:absolute;width:17px;height:17px;background:var(–zephr-color-action-main);border-radius:50%;top:2px;left:2px}.disabled.checked.svelte-17rnpjv.svelte-17rnpjv.svelte-17rnpjv:after{background:var(–zephr-color-background-tinted)}.error.svelte-17rnpjv.svelte-17rnpjv.svelte-17rnpjv{border:1px solid var(–zephr-color-warning-main)}.zephr-form-input-radio.svelte-17rnpjv:focus-visible+.zephr-form-radio-label.svelte-17rnpjv>.zephr-form-radio-dot.svelte-17rnpjv{outline:2px solid var(–zephr-color-action-main);outline-color:-webkit-focus-ring-color}
.zephr-form-input-error-text.svelte-1f2hwm0{color:var(–zephr-color-warning-main);font-family:var(–zf-root-fontFamily)}.zephr-form-input-select.svelte-1f2hwm0{display:block;appearance:auto;width:100%;height:calc(var(–zephr-input-height) * 1px);font-size:var(–zf-input-fontSize);font-family:var(–zf-root-fontFamily);border:calc(var(–zephr-input-borderWidth) * 1px) solid
var(–zephr-input-borderColor);border-radius:calc(var(–zephr-input-borderRadius) * 1px);transition:border-color 0.25s ease, box-shadow 0.25s ease;outline:0;color:var(–zephr-color-text-main);background-color:var(–zf-input-bgColor);padding:10px}.zephr-form-input-select.unselected.svelte-1f2hwm0{color:var(–zephr-color-background-tinted)}.zephr-form-input-select.error.svelte-1f2hwm0{border-color:var(–zephr-color-warning-main)}.zephr-form-input-textarea.svelte-1f2hwm0{display:block;min-height:calc(var(–zephr-input-height) * 1px);min-width:100%;max-width:100%;background-color:var(–zf-input-bgColor);border:calc(var(–zephr-input-borderWidth) * 1px) solid var(–zephr-input-borderColor);border-radius:calc(var(–zephr-input-borderRadius) * 1px);transition:border-color 0.25s ease, box-shadow 0.25s ease;color:var(–zephr-color-text-main);font-size:var(–zf-input-fontSize);font-weight:300;padding:16px 8px;font-family:var(–zf-root-fontFamily);outline:0}.zephr-form-input-textarea.svelte-1f2hwm0::placeholder{color:var(–zephr-color-background-tinted);font-family:var(–zf-root-fontFamily)}.zephr-form-input-slider-output.svelte-1f2hwm0{margin:13px 0 0 10px}
.zephr-form-link.svelte-1mdtotm{margin:10px 0;color:var(–zf-link-color);text-decoration:underline;cursor:pointer;font-family:var(–zf-root-fontFamily)}.zephr-form-link-disabled.svelte-1mdtotm{color:var(–zephr-color-text-main);cursor:not-allowed;text-decoration:none}.zephr-form-link-no-underline.svelte-1mdtotm{text-decoration:none}
.zephr-form-reset-link-container.svelte-11qigta{min-height:2.5em}.zephr-form-link-message.svelte-11qigta{margin:10px 0 10px 20px;font-family:var(–zf-root-fontFamily)}
.zephr-form-reset-link-container.svelte-11qigta{min-height:2.5em}.zephr-form-link-message.svelte-11qigta{margin:10px 0 10px 20px;font-family:var(–zf-root-fontFamily)}
.zephr-form-google-icon.svelte-155mfls{width:20px}
.zephr-form-social-sign-in.svelte-18a7qws{align-items:center}.zephr-form-social-sign-in-button.svelte-18a7qws{width:100%;min-height:55px;padding:15px;color:#000;background-color:#fff;box-shadow:0 0 5px rgba(0, 0, 0, 0.3);border-radius:10px;font-size:17px;display:flex;align-items:center;cursor:pointer;margin-top:20px;font-family:var(–zf-root-fontFamily)}.zephr-form-social-sign-in-button.svelte-18a7qws:hover{background-color:#fafafa}.zephr-form-social-sign-in-icon.svelte-18a7qws{display:flex;justify-content:center;margin-right:30px;width:25px}
.zephr-form-promo-input-container.svelte-3s0tux{flex:0.6}.zephr-form-promo-activate-line.svelte-3s0tux{display:flex;cursor:pointer;min-height:calc(var(–zephr-input-height) * 1px);margin:1em 1.2em;font-size:var(–zf-payment-summary-fontSize);line-height:calc(var(–zephr-input-height) * 1px);color:var(–zf-link-color)}.zephr-form-promo-line.svelte-3s0tux{display:flex;min-height:calc(var(–zephr-input-height) * 1px);margin:1em 1.2em}.zephr-form-addon-line.svelte-3s0tux{display:flex;padding-top:1em;padding-bottom:1em;margin-left:1.2em}.zephr-form-summary-prorate-disclaimer.svelte-3s0tux{padding-top:0.25em;padding-left:1em;color:var(–zephr-color-text-tinted)}.zephr-form-summary-charge-description.svelte-3s0tux{color:var(–zephr-color-text-tinted)}.zephr-form-charge-line.svelte-3s0tux{display:flex;padding-top:1em;margin-left:1.2em;padding-left:1em}.zephr-form-summary-line-right.svelte-3s0tux{margin-left:auto;margin-right:1.5em;text-align:right}.zephr-form-summary-line-left.svelte-3s0tux{margin-left:10px}.zephr-form-summary-promo-middle.svelte-3s0tux{position:relative;top:calc(var(–zephr-input-height) * 0.5px)}.zephr-form-summary.svelte-3s0tux{border:var(–zf-payment-summary-border);border-radius:var(–zf-payment-summary-borderRadius);padding-top:1em;padding-bottom:1em;font-size:var(–zf-payment-summary-fontSize);margin-top:1.5em;margin-bottom:1.5em;color:var(–zf-payment-summary-color);background-color:var(–zf-payment-summary-bgColor)}.zephr-form-summary.svelte-3s0tux:first-child{margin-top:0}.zephr-form-summary-title.svelte-3s0tux{margin-bottom:1em;margin-left:1.2em}.zephr-form-summary-price-line.svelte-3s0tux{font-weight:bold;padding-left:1em;padding-top:1em}.zephr-form-summary-line.svelte-3s0tux{display:flex}.zephr-form-summary-price-line-upper.svelte-3s0tux{border-top:var(–zf-payment-summary-border);margin-top:1em}.zephr-form-summary-line-title.svelte-3s0tux{flex:0 0 65%;margin-left:1.2em}.zephr-form-summary-line-price.svelte-3s0tux{flex:1;text-align:right;margin-right:1.5em}.zephr-form-discount-line.svelte-3s0tux{display:flex;justify-content:space-between;margin:0.5em 0;margin-left:1.2em}
.zephr-form-divider-container.svelte-svq2r5{display:flex;align-items:center;justify-content:center;margin:1.5em 0}.zephr-form-divider-line.svelte-svq2r5{height:1px;width:50%;margin:0 5px;background-color:var(–zephr-color-text-tinted)}.zephr-form-divider-line-complete.svelte-svq2r5{height:1px;width:100%;margin:0 5px;background-color:var(–zephr-color-text-tinted)}.zephr-form-divider-text.svelte-svq2r5{margin:0 12px;color:var(–zephr-color-text-main);font-size:var(–zf-info-fontSize);font-family:var(–zf-root-fontFamily);white-space:nowrap}
.zephr-form-login-link.svelte-1bpa3mq.svelte-1bpa3mq{margin-top:1em;margin-bottom:1em;text-align:center}.zephr-form-login-link.svelte-1bpa3mq>a.svelte-1bpa3mq{color:var(–zf-link-color);margin-left:0.25em}
.zephr-form-account-code-container.svelte-ctt9he{margin-top:1.5em}
.zephr-form-invitation-container.svelte-yy7ok3{border:1px solid;margin:1.5em 0;padding:20px;color:var(–zephr-color-text-main)}.zephr-form-invitation-title.svelte-yy7ok3{margin-bottom:20px;font-weight:bold;font-family:var(–zf-root-fontFamily)}.zephr-form-invitation-email.svelte-yy7ok3{display:flex;align-items:center}.zephr-form-invitation-input-container.svelte-yy7ok3{flex-grow:1}.zephr-form-invitation-success-message.svelte-yy7ok3{margin-top:0.5em}
.subscription-option-product-label{margin-bottom:5px;margin-top:0}.subscription-option-wrapper.svelte-1ecrab7{display:flex;flex:1 0 28%;width:100%;max-width:400px}.subscription-option-wrapper.svelte-1ecrab7:only-child{max-width:325px}.subscription-option-box{display:flex;flex-direction:column;flex:1 1 0;width:0}.info-row.svelte-1ecrab7{margin-top:7px;margin-bottom:7px}.info-label.svelte-1ecrab7{display:inline;margin-right:10px}.info.svelte-1ecrab7{display:inline}hr.svelte-1ecrab7{border-width:1px 0px 0px;margin:5px 5px;border-color:var(–zephr-color-background-tinted)}.inner-box.svelte-1ecrab7{border:var(–zf-payment-options-border);border-radius:var(–zf-payment-options-borderRadius);background-color:var(–zf-payment-options-bgColor);height:100%;display:flex;flex-direction:column;flex-grow:1;padding:20px 20px;margin:0 10px;overflow:hidden}.change-disabled-label.svelte-1ecrab7{margin-top:15px}
.subscription-options-container.svelte-1nzbr4y{flex-wrap:wrap;font-size:var(–zf-payment-options-fontSize);width:100%;color:var(–zf-payment-options-color);line-height:18px;padding:31px 0 20px;display:grid;justify-content:space-evenly;justify-items:center;grid-template-columns:repeat(auto-fit, minmax(300px, 1fr));gap:13px}@media(max-width: 340px){.subscription-options-container.svelte-1nzbr4y{grid-template-columns:repeat(auto-fit, minmax(280px, 1fr))}}
.subscription-list-item.svelte-1wg9q7{font-family:var(–zephr-typography-body-font), var(–zephr-typography-body-fallbackFont);color:var(–zephr-color-text-main)}hr.svelte-1wg9q7{border-width:1px 0px 0px;margin:20px 0px;border-color:var(–zephr-color-background-tinted)}.inactive.svelte-1wg9q7{text-decoration:line-through}h3.svelte-1wg9q7{margin:0 0 20px}.inner-box.svelte-1wg9q7{display:flex;flex-direction:column;flex-grow:1;padding:30px 20px;margin:0 10px;overflow:hidden;min-width:300px}
.zephr-form-tablink.svelte-ky9lgg{display:flex;align-items:center;cursor:pointer;justify-content:space-between;text-decoration:none;color:var(–zephr-color-text-main);font-weight:590;padding:var(–zf-list-subscriptions-section-padding);border-bottom:1px solid var(–zephr-color-background-tinted)}.zephr-form-tablink.svelte-ky9lgg:hover{text-decoration:underline}.zephr-form-tablink.svelte-ky9lgg:last-child{border-bottom:none
}
.zephr-subscription-list-box-item.svelte-h57nmp.svelte-h57nmp{font-family:var(–zephr-typography-body-font), var(–zephr-typography-body-fallbackFont);color:var(–zephr-color-text-main);max-width:var(–zf-list-subscriptions-container-maxWidth);border-radius:var(–zf-list-subscriptions-container-borderRadius);border:var(–zf-list-subscriptions-container-border);padding:var(–zf-list-subscriptions-container-padding)}.zephr-inner-box.svelte-h57nmp.svelte-h57nmp{display:flex;flex-direction:column;flex-grow:1}.zephr-box-item-information.svelte-h57nmp.svelte-h57nmp{font-size:var(–zf-list-subscriptions-info-fontSize);color:var(–zephr-color-text-tinted);margin-top:8px}.zephr-section.svelte-h57nmp.svelte-h57nmp{padding:var(–zf-list-subscriptions-section-padding);border-bottom:1px solid var(–zephr-color-background-tinted);gap:8px}.zephr-section.svelte-h57nmp.svelte-h57nmp:first-child{padding-top:0;border-bottom:none
}.zephr-section.svelte-h57nmp.svelte-h57nmp:last-child{border-bottom:none
}.zephr-section.svelte-h57nmp span.svelte-h57nmp{font-weight:590}.zephr-section.svelte-h57nmp h2.svelte-h57nmp{margin:0}
.subscriptions-list-container.svelte-2hm0a{margin-top:40px;display:flex;flex-direction:column;gap:40px}
.past-subscriptions-link-wrapper.svelte-1gacumx{margin:10px 0}
.event-link-wrapper{margin-top:10px}.float-right.svelte-12vtv9v{float:right;margin-top:calc(var(–zephr-input-height) * -1px + 5px)}
.bold-label.svelte-121zkre label{font-weight:bold}.read-only-input.svelte-121zkre{border:none;background:none;color:#333;cursor:default;outline:none;flex-grow:1;margin-right:10px;margin-top:15px}.zephr-form-relative-container.svelte-121zkre{display:flex;align-items:center;justify-content:space-between;width:100%}.update-link{white-space:nowrap;text-decoration:none;font-size:0.75em}
.bold-label.svelte-121zkre label{font-weight:bold}.read-only-input.svelte-121zkre{border:none;background:none;color:#333;cursor:default;outline:none;flex-grow:1;margin-right:10px;margin-top:15px}.zephr-form-relative-container.svelte-121zkre{display:flex;align-items:center;justify-content:space-between;width:100%}.update-link{white-space:nowrap;text-decoration:none;font-size:0.75em}
.zephr-form-ro-attribute{display:flex;align-items:center;height:50px;justify-content:flex-start}.zephr-form-ro-attribute-label{margin-right:10px;&:after {
content: “: “;
}}
hr.svelte-4jb3ht{border-width:1px 0px 0px;margin:20px 0px;border-color:var(–zephr-color-background-tinted)}.session-container.svelte-4jb3ht{padding:15px;display:flex;flex-direction:column}.device-title.svelte-4jb3ht{font-weight:bold;font-size:1.2em;margin-bottom:10px}.current-device.svelte-4jb3ht{color:#777;font-size:0.9em;margin-left:10px}.login-time.svelte-4jb3ht,.browser.svelte-4jb3ht,.country.svelte-4jb3ht{margin-bottom:10px;font-size:0.9em}
.sign-out-all.svelte-1gdzu2p{color:#0073e6;cursor:pointer;font-size:1em;margin:20px 0;text-align:center}.sign-out-all.svelte-1gdzu2p:hover{text-decoration:underline}
.table.svelte-15e6uru.svelte-15e6uru{display:grid;grid-template-columns:0.7fr 1.4fr 1fr 0.6fr}.table-head.svelte-15e6uru.svelte-15e6uru{display:contents}.table-row.svelte-15e6uru.svelte-15e6uru{display:contents}.table-head.svelte-15e6uru>div.svelte-15e6uru{font-size:13px;padding:13px 10px 10px;border-bottom:1px solid var(–zephr-color-background-tinted)}.table-row.svelte-15e6uru>div.svelte-15e6uru{font-size:var(–zf-info-fontSize);height:60px;padding:0 10px;border-bottom:1px solid var(–zephr-color-background-tinted);display:flex;align-items:center}.table-head.svelte-15e6uru>div.svelte-15e6uru:first-child,.table-row.svelte-15e6uru>div.svelte-15e6uru:first-child{padding-left:0}.table-head.svelte-15e6uru>div.svelte-15e6uru:last-child,.table-row.svelte-15e6uru>div.svelte-15e6uru:last-child{padding-right:0}.taxes.svelte-15e6uru.svelte-15e6uru{color:var(–zephr-color-background-tinted);white-space:pre;font-size:small}
.zephr-form-button-group.svelte-s116f2{display:flex;flex-direction:row;gap:12px;margin-top:20px}.zephr-form-back-button.svelte-s116f2{padding:0 20px;border-radius:calc(var(–zephr-button-borderRadius) * 1px);border-style:solid;border-width:calc(var(–zephr-button-borderWidth) * 1px);border-color:var(–zephr-color-action-main);display:flex;justify-content:center;align-items:center;width:100%;background-color:var(–zephr-color-action-main);color:var(–zf-button-color);cursor:pointer}.zephr-form-continue-button.svelte-s116f2{padding:0 20px;border-radius:calc(var(–zephr-button-borderRadius) * 1px);border-style:solid;border-width:calc(var(–zephr-button-borderWidth) * 1px);border-color:var(–zephr-color-action-main);display:flex;justify-content:center;align-items:center;width:100%;background-color:var(–zf-button-color);color:var(–zephr-color-action-main);cursor:pointer}.zephr-form-back-button.svelte-s116f2:hover{background-color:var(–zephr-color-action-tinted);border-color:var(–zephr-color-action-tinted)}.zephr-form-continue-button.svelte-s116f2:hover{background-color:var(–zephr-color-action-tinted);border-color:var(–zephr-color-action-tinted);color:var(–zf-button-color)}.zephr-form-back-button.svelte-s116f2:not(inputHeight){height:calc(var(–zephr-button-height) * 1px)}.zephr-form-continue-button.svelte-s116f2:not(inputHeight){height:calc(var(–zephr-button-height) * 1px)}
.zephr-form-button.svelte-16hrghb{padding:0 20px;text-decoration:none;text-transform:capitalize;border-radius:calc(var(–zephr-button-borderRadius) * 1px);font-size:var(–zf-button-fontSize);font-weight:normal;cursor:pointer;border-style:solid;border-width:calc(var(–zephr-button-borderWidth) * 1px);border-color:var(–zephr-color-action-tinted);transition:backdrop-filter 0.2s, background-color 0.2s;display:flex;justify-content:center;align-items:center;width:100%;background-color:var(–zephr-color-action-main);color:var(–zf-button-color);position:relative;overflow:hidden;font-family:var(–zf-root-fontFamily);height:calc(var(–zephr-button-height) * 1px)}.zephr-form-button.svelte-16hrghb:hover{background-color:var(–zephr-color-action-tinted);border-color:var(–zephr-color-action-tinted);color:var(–zf-button-color)}.zephr-form-button-top-margin.svelte-16hrghb{margin-top:20px}
.recipient-name-section.svelte-wt5yoy{margin-top:1rem;margin-bottom:1rem}.field.svelte-wt5yoy{margin-bottom:0.5rem}label.svelte-wt5yoy{display:block;margin-bottom:0.5rem;color:var(–zephr-color-text-tinted);font-family:var(–zf-root-fontFamily)}input.svelte-wt5yoy{width:100%;padding:0.5rem;border:calc(var(–zephr-input-borderWidth) * 1px) solid var(–zephr-input-borderColor);border-radius:calc(var(–zephr-input-borderRadius) * 1px);background-color:var(–zf-input-bgColor);color:var(–zephr-color-text-main);font-family:var(–zf-root-fontFamily)}input.svelte-wt5yoy::placeholder{color:var(–zephr-color-background-tinted);font-family:var(–zf-root-fontFamily)}input.svelte-wt5yoy:focus{outline:none;border-color:var(–zephr-color-action-main)}.error.svelte-wt5yoy{border-color:var(–zephr-color-warning-main) !important}.error-text.svelte-wt5yoy{color:var(–zephr-color-warning-main);font-size:0.875em;margin-top:0.25rem;font-family:var(–zf-root-fontFamily)}
.recipient-email-section.svelte-17rjebp{margin-bottom:1rem}.field.svelte-17rjebp{margin-bottom:0.5rem}label.svelte-17rjebp{display:block;margin-bottom:0.5rem;color:var(–zephr-color-text-tinted);font-family:var(–zf-root-fontFamily)}input.svelte-17rjebp{width:100%;padding:0.5rem;border:calc(var(–zephr-input-borderWidth) * 1px) solid var(–zephr-input-borderColor);border-radius:calc(var(–zephr-input-borderRadius) * 1px);background-color:var(–zf-input-bgColor);color:var(–zephr-color-text-main);font-family:var(–zf-root-fontFamily)}input.svelte-17rjebp::placeholder{color:var(–zephr-color-background-tinted);font-family:var(–zf-root-fontFamily)}input.svelte-17rjebp:focus{outline:none;border-color:var(–zephr-color-action-main)}.error.svelte-17rjebp{border-color:var(–zephr-color-warning-main) !important}.error-text.svelte-17rjebp{color:var(–zephr-color-warning-main);font-size:0.875em;margin-top:0.25rem;font-family:var(–zf-root-fontFamily)}
.recipient-address-section.svelte-hqngij{margin-bottom:1rem}.address-main-label.svelte-hqngij{display:block;margin-bottom:0.75rem;color:var(–zephr-color-text-tinted);font-family:var(–zf-root-fontFamily);font-weight:600}.address-fields.svelte-hqngij{display:grid;grid-template-columns:repeat(2, 1fr);gap:1rem}.full-width.svelte-hqngij{grid-column:1 / -1}.field.svelte-hqngij{margin-bottom:0.5rem}label.svelte-hqngij{display:block;margin-bottom:0.5rem;color:var(–zephr-color-text-tinted);font-family:var(–zf-root-fontFamily)}input.svelte-hqngij{width:100%;padding:0.5rem;border:calc(var(–zephr-input-borderWidth) * 1px) solid var(–zephr-input-borderColor);border-radius:calc(var(–zephr-input-borderRadius) * 1px);background-color:var(–zf-input-bgColor);color:var(–zephr-color-text-main);font-family:var(–zf-root-fontFamily)}input.svelte-hqngij::placeholder{color:var(–zephr-color-background-tinted);font-family:var(–zf-root-fontFamily)}input.svelte-hqngij:focus{outline:none;border-color:var(–zephr-color-action-main)}.error-text.svelte-hqngij{color:var(–zephr-color-warning-main);font-size:0.875em;margin-top:0.25rem;font-family:var(–zf-root-fontFamily)}@media(max-width: 640px){.address-fields.svelte-hqngij{grid-template-columns:1fr}}
.recipient-message-section.svelte-90amkz{margin-top:1rem;margin-bottom:1rem}.field.svelte-90amkz{margin-bottom:0.5rem}label.svelte-90amkz{display:block;margin-bottom:0.5rem;color:var(–zephr-color-text-tinted);font-family:var(–zf-root-fontFamily)}textarea.svelte-90amkz{width:100%;min-height:100px;resize:vertical;padding:0.5rem;border:calc(var(–zephr-input-borderWidth) * 1px) solid var(–zephr-input-borderColor);border-radius:calc(var(–zephr-input-borderRadius) * 1px);background-color:var(–zf-input-bgColor);color:var(–zephr-color-text-main);font-family:var(–zf-root-fontFamily)}textarea.svelte-90amkz::placeholder{color:var(–zephr-color-background-tinted);font-family:var(–zf-root-fontFamily)}textarea.svelte-90amkz:focus{outline:none;border-color:var(–zephr-color-action-main)}
.zephr-form-input-wrapper.svelte-1wfl3xs{display:flex;flex-direction:column;margin-bottom:16px;width:100%}label.svelte-1wfl3xs{margin-bottom:8px;font-weight:500}input.svelte-1wfl3xs{padding:12px;border-radius:var(–input-border-radius, 4px);border:var(–input-border-width, 1px) solid var(–input-border-color, #ccc);font-size:16px;height:var(–input-height, 50px);box-sizing:border-box}input.svelte-1wfl3xs:focus{outline:none;border-color:var(–color-action-main, #006EDB)}input.error.svelte-1wfl3xs{border-color:var(–color-warning-main, #D90B00)}.error-message.svelte-1wfl3xs{color:var(–color-warning-main, #D90B00);font-size:14px;margin-top:4px}.required.svelte-1wfl3xs{color:var(–color-warning-main, #D90B00)}
.zephr-form-max-width{max-width:var(–zf-container-maxWidth);margin:auto}.zephr-form-content.svelte-cde0t0{padding:var(–zf-container-padding);background-color:var(–zf-container-bgColor);border:var(–zf-container-border);border-radius:var(–zf-container-borderRadius)}.zephr-payment-options-content.svelte-cde0t0{background-color:transparent;border:none;border-radius:0}.close-button-container.svelte-cde0t0{text-align:right;width:100%;margin-bottom:15px}.close-button.svelte-cde0t0{cursor:pointer;display:inline;padding:7px;min-width:32px;min-height:32px;border-radius:16px;margin:1px}.loading-placeholder.svelte-cde0t0{display:flex;justify-content:center;align-items:center;min-height:200px}
.zephr-form{–zf-root-fontSize:16px;–zf-root-fontFamily:var(–zephr-typography-body-font), var(–zephr-typography-body-fallbackFont);–zf-container-maxWidth:440px;–zf-container-padding:20px;–zf-container-margin:20px auto;–zf-container-border:none;–zf-container-borderRadius:calc(var(–zephr-input-borderRadius) * 1px);–zf-container-color:var(–zephr-color-text-main);–zf-container-bgColor:var(–zephr-color-background-main, transparent);–zf-info-fontSize:14px;–zf-subtext-fontSize:12px;–zf-link-color:#6ba5e9;–zf-input-fontSize:var(–zf-root-fontSize);–zf-input-bgColor:var(–zephr-color-background-main, transparent);–zf-button-fontSize:calc(var(–zephr-button-fontSize, 16) * 1px);–zf-button-color:#fff;–zf-password-valid-color:#7bcb7f;–zf-password-invalid-color:#c4c4c4;–zf-payment-options-fontSize:var(–zf-subtext-fontSize);–zf-payment-options-color:var(–zephr-color-text-main);–zf-payment-options-bgColor:var(–zf-container-bgColor);–zf-payment-options-border:calc(var(–zephr-input-borderWidth) * 1px) solid var(–zephr-input-borderColor);–zf-payment-options-borderRadius:calc(var(–zephr-input-borderRadius) * 1px);–zf-payment-options-highlighted-color:var(–zephr-color-text-main);–zf-payment-options-highlighted-bgColor:var(–zephr-color-accent-main);–zf-payment-summary-fontSize:var(–zf-subtext-fontSize);–zf-payment-summary-color:var(–zephr-color-text-main);–zf-payment-summary-bgColor:var(–zf-input-bgColor);–zf-payment-summary-border:calc(var(–zephr-input-borderWidth) * 1px) solid var(–zephr-input-borderColor);–zf-payment-summary-borderRadius:calc(var(–zephr-input-borderRadius) * 1px);–zf-list-subscriptions-container-maxWidth:630px;–zf-list-subscriptions-container-padding:24px;–zf-list-subscriptions-section-padding:16px 0px;–zf-list-subscriptions-container-border:calc(var(–zephr-input-borderWidth) * 1px) solid var(–zephr-input-borderColor);–zf-list-subscriptions-container-borderRadius:calc(var(–zephr-input-borderRadius) * 1px);–zf-list-subscriptions-info-fontSize:var(–zf-info-fontSize)}.zephr-form.svelte-jjl1p8{color:var(–zf-container-color);margin:var(–zf-container-margin);font-size:var(–zf-root-fontSize);font-family:var(–zf-root-fontFamily)}.zephr-form *{box-sizing:border-box}:where(.zephr-form) button{text-transform:unset;overflow:visible;font-family:var(–zf-root-fontFamily, unset);font-size:var(–zf-root-fontSize, unset);line-height:unset;margin:unset;padding:unset;border:unset;background:unset;&:hover, &:focus-visible {
border: unset;
background: unset;
}}.zephr-form-disabled-div{pointer-events:none;opacity:0.8}.zephr-form-relative-container{position:relative}.zephr-form-flex-container{display:flex}.zephr-form-input-inner-button{cursor:pointer;position:absolute;top:50%;transform:translateY(-50%);right:5px;padding:10px 5px}.zephr-form-input-inner-text{color:var(–zephr-color-text-main);font-size:var(–zf-subtext-fontSize);font-weight:bold;font-family:var(–zf-root-fontFamily)}
]]>
The stakes are growing higher as AI continues to operate more independently. During internal cybersecurity evaluations, OpenAI models circumvented controls designed to isolate them from the internet, compromising parts of OpenAI’s internal research infrastructure and Hugging Face’s systems. Episodes like that have raised interest in ways to catch suspicious behavior before an AI agent can act on it. OpenAI says it is now investing more computing resources in chain-of-thought monitoring.
But machine learning engineer Shikhar Shiromani and AI researcher Leo Richter’s new work suggests that this type of monitoring may have a weakness. In a subset of 251 cases in which reasoning was the only clue to the problem, the monitor’s detection rate fell from 96.2 percent to 3.8 percent after the researchers rewrote the reasoning to make the behavior appear innocent without changing its actions.
“The AI isn’t necessarily lying about its action,” says Shiromani, a research fellow at the nonprofit Pivotal. “It is more or less giving a misleading, innocent-sounding explanation of its intentions.”
Jacob Andreas, a computer scientist at MIT, says that chain-of-thought, or CoT, reasoning can provide clues about what a model is trying to do.
“But we should be skeptical: (a) that any individual CoT provides us insight into model behavior on a specific example, and (b) that absence of evidence of bad behavior in a CoT should be taken as evidence of absence,” he wrote in an email.
Andreas questioned how much Shiromani’s main experiment says about real-world monitoring. Because the researchers rewrote the reasoning after the fact, he says, it is unclear whether a real model could produce the same innocent-sounding reasoning while taking the suspicious action, such as trying to hack another system.
The researchers also tested live AI agents. After adjusting the monitor for each model, it caught every GPT-5.4 hack and all but one DeepSeek hack. Claude Opus was harder to catch, with the monitor missing seven of 20 hacks on one task and 12 of 20 on another.
“And at the end of the day, there is no substitute now for really rigorous behavioral testing, and no substitute for human operators simply declining to run agents in situations where there is potential for substantial harm,” Andreas says.
